Local models and model tooling

Inference, quantization, hardware, fine-tuning, LoRAs, evaluation and deployment

• articleWhat does the sd-scripts FLUX.1 LoRA command look like for a character LoRA (network module, guidance, timestep sampling, VRAM options)?
  Problem. FLUX.1 differs from SD, so SD-style LoRA settings do not transfer one-to-one. Conditions. flux_train_network.py in sd-scripts with FLUX.1 weights, CLIP-L, T5-XXL and AE files. Method. The documented example uses --network_module=networks.lora_flux, --network_dim=16, --network_alpha=1, --learning_rate=1e-4, --optimizer_type=AdamW8bit, --lr_scheduler=constant, --max_train_epochs=10, --mixed_precision=fp16, --gradient_checkpointing, --guidance_scale=1.0, --timestep_sampling=flux_shift, --model_prediction_type=raw, --blocks_to_swap=18, --cache_text_encoder_outputs, --cache_latents. For Chroma, --guidance_scale=0.0 is required. Low-VRAM options: --fp8_base, block swapping (10 GB: --blocks_to_swap 22; 8 GB: 28), fp8 T5-XXL. Dataset and caption keys are shared with the SD guide (see C3-B571-07). Failure modes. fp8 training results 'may vary' per the doc. Limitations. Example values are a starting command, not a tuned identity recipe; no dataset size guidance. Version/date. flux_tr...
• articleHow do I run SAM 2's video predictor to get per-object masklets for compositing plates (init_state, prompts, propagate_in_video)?
  Problem. Isolating a generated subject across frames for compositing. Conditions. Python >= 3.10, torch >= 2.5.1, torchvision >= 0.20.1, CUDA toolkit matching PyTorch for the custom kernel. Method. build_sam2_video_predictor(model_cfg, checkpoint) with e.g. configs/sam2.1/sam2.1_hiera_l.yaml and sam2.1_hiera_large.pt; inside torch.inference_mode() with autocast bfloat16: state = predictor.init_state(video); predictor.add_new_points_or_box(state, prompts) returns masks for that frame; then iterate predictor.propagate_in_video(state) to get masklets for all frames. The 12/11/2024 update adds vos_optimized=True (torch.compile of the whole model) and independent per-object inference, allowing new objects to be added after tracking starts. Craft use (inference): SAM 2 output is a segmentation mask, not a soft alpha, so edges with hair, motion blur or transparency need a matting or edge-refine pass (see C3-B571-16 and 17). Failure modes. See the existing record for SAM 2's reported failur...
• articleHow do I set Robust Video Matting's downsample_ratio and recurrent state for generated human footage, and what is its licence?
  Problem. Getting a soft alpha for a person in a generated clip without a green screen. Conditions. Human subjects; RVM PyTorch/TorchScript models. Method. (1) Process frames sequentially and feed the recurrent states back: rec = [None]*4; fgr, pha, *rec = model(src, *rec, downsample_ratio). (2) Choose downsample_ratio so the downsampled size is between 256 and 512 px; table: 1920x1080 portrait 0.25, full body 0.4; 1280x720 0.375 / 0.6; 3840x2160 0.125 / 0.2; <=512x512 use 1. Higher is not always better. (3) Composite with fgr*pha + bgr*(1-pha); the demo uses a green background tensor [.47, 1, .6]. (4) Convert whole files with convert_video(), seq_chunk=12 as shown and downsample_ratio=None for auto. Failure modes. Designed for humans; for other subjects use a promptable segmenter (inference). Limitations. Code is GPL-3.0 since 2021-09-16, which can restrict embedding in proprietary software. Version/date. README and inference doc fetched 2026-10-23.
• articleHow do I use RIFE and FILM to raise frame rate of an AI clip (exp, scale, times_to_interpolate), and what do they not fix?
  Problem. Generated clips are often short and low-fps; interpolation adds frames but does not repair identity changes. Conditions. Local RIFE (MIT) or FILM (Apache-2.0). Method. RIFE: python3 inference_video.py --exp=1 --video=video.mp4 (exp=1 gives 2X; exp=2 gives 4X, so exp = log2 of the factor), optional --fps; for very high resolution such as 4K set --scale=0.5; if disordered patterns appear try --scale=2.0 (this controls the processing resolution for the optical flow model). FILM: --times_to_interpolate N yields (2^N + 1)*(num_frames - 1) output frames and the tool targets near-duplicate frames with large motion; frames must sort naturally by name. Inference: for 24 to 48 fps use RIFE exp=1 or FILM times_to_interpolate=1 per frame pair; interpolation synthesises in-betweens from two neighbours, so it cannot remove a face change that already exists between them. Failure modes. RIFE flow errors on fast motion show as 'disordered pattern' (scale parameter); see the cycle's existing...
• articleFor a character LoRA with kohya sd-scripts, which dataset-config keys control repeats and captions, what do the docs say about rank, alpha and learning rate, and how do I watch for overfitting?
  Problem. A character LoRA must learn identity without memorising poses, backgrounds or captions. Conditions. sd-scripts train_network.py (SD 1.x/2.x) with a dataset .toml. Method. (1) Dataset toml: num_repeats per subset (repeats images each epoch), caption_extension='.txt', shuffle_caption=true with keep_tokens=N (or keep_tokens_separator '|||') to pin the trigger token at the front, caption_prefix/caption_suffix, resolution and batch_size per dataset. (2) Network: --network_module=networks.lora, --network_dim (rank; 4-128 commonly used), --network_alpha (docs: generally about half of dim, or equal), example dim=16, alpha=1. (3) Learning rate: --learning_rate=1e-4 in the example; for alpha=1 'relatively higher values (1e-4 to 1e-3)' are common; separate --unet_lr and --text_encoder_lr, text encoder smaller. (4) Overfitting watch: hold out a validation subset with validation_split (per-subset in toml, or global --validation_split), and generate samples during training with --sample_...
• articleFor speech and video datasets, how often does the dataset licence disagree with the terms of the underlying sources, and where can I see the per-dataset records?
  Corrected 2026-10-23 - see C3-B643-05; the original statements below are kept for history and are superseded where the correction says so. The Data Provenance Initiative's 2025 audit (arXiv 2412.17847v2) reports, for 3916 datasets, that 25%, 33% and 32% of text, speech and video datasets carry non-commercial licences but that most of their source content carries non-commercial restrictions through source terms (author-reported); its released AMI record lists the dataset licence as CC BY 4.0 while the GitHub licence inferred from a converter repository (AMICorpusXML) is MIT, and Papers with Code reports Unspecified.
• articleWhat must an agent record alongside a ComfyUI workflow so it can be re-run faithfully later?
  The two JSON forms leave out most of what decides the output. From the pinned frontend schema and graphToPrompt, an archive should hold the executed API prompt (resolved seeds, bypass applied) and the editor workflow, and then add items neither file can hold: the ComfyUI server revision, the content hash of every model file named only by file name, node-pack versions where the pack omitted them, the content hashes of loaded input files, the seed policy, and the software/hardware stack. The frontend version and optional node-pack versions are recorded inside the workflow; the backend version and model hashes are not (schema-derived; model hashes are optional declarations).
• articleWhich results from planning-with-a-learned-model papers could transfer to media or agent workflows, and which are specific to robotics or games?
  In the inspected papers the demonstrated uses are games, simulated control and robot manipulation; none reports a media-production or LLM-agent workflow. What may transfer is method-level (plan by sampling sequences through a learned predictor, check model exploitation by comparing virtual versus real score, use a latent rather than pixel predictor for speed) not the results. Specifically, V-JEPA 2-AC needed a goal image and a reward-free robot dataset; Dreamer and TD-MPC2 need reward signals; the exploit-the-dream failure from 2018 is the most portable lesson. This is a synthesis; no experiment was run.
• articleWhich training datasets does each of Real-ESRGAN, FILM, RIFE and Practical-RIFE itself name, and where is that stated?
  Real-ESRGAN names DIV2K, Flickr2K and OutdoorSceneTraining (OST) in its paper and DF2K plus OST in docs/Training.md. FILM names Vimeo-90K as its main training dataset (README) and as training data (paper). RIFE names Vimeo90K in its paper and README. The Practical-RIFE README at the pinned commit names no training dataset, so its data lineage is unknown from that file. REDS is not named in the four model repositories read.
• articleWhat terms and scope do the public TD-MPC2 code, checkpoints and datasets state, and what is left open?
  TD-MPC2 has the clearest statements of the three planning systems: the repository README says the project is MIT-licensed (LICENSE: MIT, Copyright Nicklas Hansen 2023) with third-party code under its own licences; the Hugging Face model card says in prose '**License:** MIT' for the 324 checkpoints (12 multi-task from 1M to 317M, 312 single-task at 5M), and the dataset card says MIT for the 30-task and 80-task replay-buffer datasets (545M and 345M transitions). The card's out-of-scope note says checkpoints are not expected to generalise to unseen tasks. Not stated: redistribution of simulator-derived data, model_training on outputs, and the terms of the underlying simulators (DMControl, Meta-World, ManiSkill2, MyoSuite) and of MuJoCo assets.
• articleWhat actually helped in an 8-day AI video-editing trial (Jun 2026): editor plugins, agent-made B-roll, Claude + HyperFrames motion graphics, Omni footage edits
  Creator-reported: statements attributed to the creator(s) come from discovery's reading of auto-caption transcript excerpts; the curator checked the video metadata (title, date, length) and any cited documentation pages, not the spoken claims. Nothing was tested; promotional or sponsored framing applies where stated. Dan Kieft and team (2026-06-12) sorted tools into money-saving, speed and creativity. Useful: Higgsfield plugins inside Premiere Pro, After Effects and Figma (generate shots, reframe 16:9 to vertical, 4K upscale; background removal "mixed results"); Higgsfield "Supercomputer" agent producing intro scenes and B-roll/infographic "cards" for explainers, with failures retried on request; Claude Code + HyperFrames (HeyGen's Apache-2.0 open-source HTML-to-video renderer "built for agents", R6) to add titles, captions, lists, counters and UI animations around a raw talking-head recording by describing them, then iterating in plain language ("make the transition faster"); Googl...
• articleDirecting AI camera moves with a Blender previs clip: an agent builds the 3D blockout and camera, then the render is fed to Seedance as a "camera movement reference only"
  Creator-reported: statements attributed to the creator(s) come from discovery's reading of auto-caption transcript excerpts; the curator checked the video metadata (title, date, length) and any cited documentation pages, not the spoken claims. Nothing was tested; promotional or sponsored framing applies where stated. Dan Kieft (2026-08-05). Problem: complex camera choreography is hard to describe, and re-prompting never reproduces a lucky shot. Method: (1) make a cinematic start frame (any image model); (2) give the frame and a written camera plan (e.g. 15 s: push in on one person, move to another, pass through a window to reveal a food stand) to Claude connected to Blender via MCP - Claude builds a rough scene and animated camera (took about 5+ minutes on his PC); preview, then Render > Render Animation; (3) in Seedance 2.0/2.5 attach the start frame plus the Blender render and begin the prompt with "video 1 as a camera movement reference only", then describe action/dialogue. His A...
• articleWhen should I train a style LoRA instead of using reference images, and what do BFL's FLUX.2 klein guide say about dataset size, captions and trigger words?
  Problem. A brand look must hold across hundreds of images, offline or at low cost. Conditions. FLUX.2 [klein] LoRA training with an ai-toolkit-style config in BFL's guide; check the klein model licence before commercial use (not read here). Method (documented). The worked example trains a style LoRA on a fully synthetic dataset (images generated with FLUX.1 [dev] and illustrative LoRAs): 27 training images with caption files. Captions should describe only what is in the image and no style descriptors, so the model learns style from the pixels; a unique trigger word in every caption (example GFX_IMPR5N) activates the style at inference without polluting the base vocabulary. Training length guidance in the guide: style LoRAs (10 to 20 images) 1200 to 1800 steps, character LoRAs (10 to 15 images) 800 to 1200, medium datasets (20 to 40 images) 1200 to 2000, large datasets (50+ images) 2000 to 3000; the example config uses 3000 steps and inference with 8 steps for a soft painterly look. ...
• articleWhat do open AI video tools (Wan2.2 save_video, diffusers export_to_video) actually write to the file: bit depth, encoder, colour tags, resolution?
  Problem. A generated clip looks different after export (banding, size change, shifted levels) and nobody knows what the tool did. Conditions. Python tools that write MP4 through imageio/ffmpeg. Method (read from code). Wan2.2 wan/utils/utils.py save_video: tensor clamped to value_range (-1,1), made into a grid per frame, multiplied by 255 and cast with .type(torch.uint8), written with imageio.get_writer(codec='libx264', quality=8); fps default 30 in the signature. diffusers export_utils.export_to_video: float numpy frames become (frame*255).astype(np.uint8); imageio.get_writer with quality default 5 (variable bit rate; 10 highest), optional bitrate, and macro_block_size default 16, where imageio tells ffmpeg to scale the image up to the next size divisible by that number; set None or 1 to disable, with the warning that many players cannot decode odd sizes. Observations: (a) 8-bit output, so banding risk (C3-B586-01/02); (b) float to uint8 is a cast, not documented as rounded or dith...
• articleWhat do the VibeVoice authors say about long-form multi-speaker generation: length, speaker count, evaluation, limitations and misuse risks?
  Problem: choose a model for podcast or multi-participant audio. Conditions: dialogue transcripts with speaker labels. Method (VibeVoice technical report): a 7.5 Hz continuous speech tokenizer plus an LLM with next-token diffusion; up to 90 minutes in a 64K context window with a maximum of four speakers. Evaluation (authors'): 8 long transcripts totalling about 1 hour; WER via Whisper-large-v3 and Nemo, speaker similarity via WavLM, and MOS from 24 annotators on realism, richness and preference, each listening to about six hours; comparisons include ElevenLabs v3 alpha and Gemini 2.5 Pro TTS. Stated limits: English and Chinese only; no non-speech audio (background noise, music, sound effects); no overlapping speech; deepfake and disinformation risk; the authors do not recommend commercial or real-world use without further testing and describe the model as research and development only. Failure modes: unexpected audio for other languages; long output drift is not separately measured i...
• articleHow does Sesame's CSM use conversation context for speaker consistency, and what are its documented limits and misuse prohibitions?
  Problem: keep two speakers consistent across turns in generated conversation. Conditions: CSM-1B base model on a CUDA GPU, with Llama-3.2-1B and CSM-1B access. Method (README): CSM generates RVQ audio codes from text and audio inputs using a Llama backbone and a small audio decoder producing Mimi codes. With no prompt it uses a random speaker identity; to keep voices consistent, pass previous turns as Segment objects (text, speaker id, audio) as context, because the README says CSM sounds best when provided with context. The example call sets max_audio_length_ms=10_000, so long outputs have to be built turn by turn. The release is a base model, not fine-tuned to any voice, and it generates audio only (an LLM must write the text). Non-English is weak despite some capacity. Misuse: the README explicitly prohibits impersonation or fraud (mimicking real individuals without explicit consent), misinformation or deception, and illegal or harmful use. Failure modes: voice drift when context...
• articleThe diffusers memory guide says the Wan VAE does not support tiling or slicing, but the class has enable_tiling: how can a newcomer tell which statement is out of date without running anything?
  Partially verified: the curator re-read none of the cited sources; S1, S2, S6, S4, S12, S13, S14, S16, S17 rest on the producer's pinned reading. A newcomer can compare dates: in diffusers, the two Wan warnings in the memory guide were added on 2025-05-01, and the Wan VAE tiling/slicing code was merged on 2025-05-05, so the code is newer than the warnings; the guide's tiling sentence has not changed since. The class defines enable_tiling itself and gets enable_slicing from a shared mixin, and slicing only helps when decoding more than one video at once. Treat the warning as likely stale but unconfirmed until a short run on your own hardware, because several merged bug-fix PRs show Wan2.2 tiling had errors that were later fixed and two related PRs are still open.
• articleMy GPU cannot hold a video pipeline in diffusers: which memory-saving calls exist, what does each one cost, and what do the Wan, LTX-Video, HunyuanVideo and CogVideoX pages tell me to try first?
  Partially verified: the curator re-read none of the cited sources; S1, S14, S2, S8, S9, S10, S11, S25, S27, S26, S28, S34, S35 rest on the producer's pinned reading. At diffusers v0.41.0 (tag commit 086bf9578c0f4acbc66e48cd1e7cc26befd9e10f) the docs and docstrings describe three offloading levels: enable_model_cpu_offload (whole components move; smaller saving, much faster), enable_sequential_cpu_offload (submodule level; largest saving, 'extremely slow') and group offloading (block_level or leaf_level; stated to sit between the two). Each of the four video pipeline pages gives one example with a stated approximate VRAM figure (Wan ~13GB, LTX-Video ~10GB, HunyuanVideo ~14GB, CogVideoX-5B ~16GB), but each example combines different options, no sentence names a GPU or a peak-memory metric, and only the LTX example sets its own resolution (Wan and HunyuanVideo use pipeline defaults for size), so the four numbers are not comparable and are not a feasibility test for a given machine. No ...
• articleDo the Objaverse-trained 4D generators (L4GM, SV4D, SV4D 2.0, Diffusion4D) say whether their training data overlaps the Consistent4D benchmark objects, and what should an agent assume when comparing their scores?
  Partially verified: the curator re-read none of the cited sources; S14, S13, S9, S5, S15, S16, S8 rest on the producer's pinned reading. The Consistent4D author states that six of the seven benchmark objects are in Objaverse and recommends filtering them from training. A set intersection I ran (synthesis, not executed code from the repositories) finds 6 of the 7 benchmark uids in a 37,789-ID list of animated Objaverse-1.0 objects published in the Diffusion4D repository, matching that statement. SV4D states it excluded the benchmark objects; L4GM v1 states that it renders all animated Objaverse 1.0 objects and evaluates on the benchmark, and the passages searched contain no exclusion statement; SV4D 2.0 v3 contains no such statement in the strings searched. Overlap is therefore an open question for L4GM, and possibly SV4D 2.0, on this benchmark, not an established fact: L4GM filters videos by motion, and the L4GM training list is not in its repository.
• articleWhat does the safetensors format define (header layout, __metadata__, dtypes, alignment) and how can an agent read a model file's header without loading weights?
  Corrected 2026-10-23 - see C3-B595-05; the original statements below are kept for history and are superseded where the correction says so. A .safetensors file is: 8 bytes (little-endian u64 N), N bytes of UTF-8 JSON header, then a raw byte buffer. The header maps tensor names to {dtype, shape, data_offsets[begin,end]} (offsets relative to the start of the byte buffer) and may contain a '__metadata__' string-to-string map. At the inspected commit the library accepts the dtypes listed in tensor.rs (BOOL through F64/U64, BF16, several FP8 and sub-byte types), caps the header at 100,000,000 bytes, requires contiguous gap-free offsets, and its writer sorts tensors by dtype alignment and pads the header with spaces to a multiple of 8. An agent can read the header with a few lines of standard code (no ML library) and learn tensor names, shapes, dtypes and metadata.
• articleHow many frames, views and what resolution does each text/image/video-to-4D system produce, and what GPU time and memory do the authors report?
  Corrected 2026-10-23 - see C3-B588-07; the original statements below are kept for history and are superseded where the correction says so. Partially verified: the curator re-read none of the cited sources; S1, S11, S3, S4, S13, S14, S15, S16, S17, S5, S6, S8 rest on the producer's pinned reading. Reported per-asset cost spans about four to five orders of magnitude, all author-measured on different hardware. Seconds: L4GM (0.3 s per 16-frame pass plus about 2 s for views, RTX 4080 Super). Minutes: DreamGaussian4D (6.5-10 min, 80 GB A100), Diffusion4D (8 min, paper), SV4D (15-20 min fitting) and SV4D 2.0 (1 min generation + 21 min fitting, up to 30 GB, H100). Hours: Consistent4D (about 2-2.5 h, V100), MAV3D (about 6.5 h on 8 A100) and 4D-fy (about 23 h on one 80 GB A100). Clips are short: 12-32 frames per pass depending on the system (12, 14, 16, 21, 24, 32), time axes are continuous for the NeRF/neural-field systems, and no passage read states output duration in seconds for text-to-4...
• articleWhich of the driving and robot world models can an agent actually download and run, on what hardware, and what does each component's text say?
  Corrected 2026-10-23 - see C3-B598-04; the original statements below are kept for history and are superseded where the correction says so. Open and runnable: Vista (weights on Hugging Face, 32 GB VRAM for sampling), Cosmos-Predict1 driving multi-view checkpoints (three types in the downloader) and Cosmos-Predict2.5 driving multiview and robot action-conditioned models (README/table listed), and the 1X challenge GENIE baselines. Code only: DriveDreamer (research code; no weights named in README/DOCS). Not found: GAIA-1/2/3/4 (no repository, release or access statement found) and 1XWM (no release stated). Hardware is large for the Cosmos multiview models (8 GPUs of 80 GB for Predict2.5 multiview inference). Rights differ per component and there is one same-work tension: Vista's card says Apache-2.0 while its paper says it was initialised from SVD, whose checkpoint the authors call non-commercial. Cosmos licence clauses are covered by C3-B019-W6 and are not repeated.
• articleWhich open image-to-3D and text-to-3D asset generators have released code and weights, and what file or representation does each one produce?
  Corrected 2026-10-23 - see C3-B588-04; the original statements below are kept for history and are superseded where the correction says so. Eight open releases were read at pinned commits. Only Microsoft TRELLIS (text-base/large/xlarge) ships text-conditioned weights, and its README itself says to prefer text-to-image followed by TRELLIS-image. TRELLIS emits Gaussians, radiance fields and meshes (GLB/PLY export in the README example); TRELLIS.2 emits a mesh with PBR maps (GLB); Hunyuan3D 2.x emits a trimesh shape then a separate texturing stage (2.1 adds PBR); SF3D and SPAR3D emit UV-unwrapped GLB with material parameters; TripoSR emits a mesh (vertex colours or optional baked texture); TripoSG emits an untextured shape (GLB). Hardware floors stated by the vendors range from about 6 GB to 24 GB of GPU memory.
• articleWhich model file formats can run code when loaded in ComfyUI, and what protections exist at the pinned versions?
  Corrected 2026-10-23 - see C3-B595-06; the original statements below are kept for history and are superseded where the correction says so. Pickle-based model files (.ckpt, .pt, .pth, .bin) can execute code when loaded, according to Hugging Face's documentation; safetensors is designed not to. ComfyUI at commit 46aad33e loads .safetensors/.sft files with safetensors.safe_open and everything else with torch.load(..., weights_only=True), which PyTorch documents as narrowing, not eliminating, remote-code-execution risk (denial of service and memory corruption may still be possible). ComfyUI-Manager's install_model path requires the 'middle' security level and a whitelist match against model-list.json, and for non-.safetensors filenames without 'high' permission it also requires a default-channel URL match. The ComfyUI frontend's missing-model download allowlist (huggingface.co, civitai.com and civitai.red prefixes plus a few exact URLs) still permits .ckpt, .pt and .pth suffixes. Huggin...
• articleWhich primary pages support the Vista (32 GB, Apache-2.0) and Cosmos-Predict2.5 multiview (8x80 GB) hardware and licence statements previously cited only to the Apache licence text?
  Correction to C3-B144-W29-05 (nature: citation does not support the claims; facts hold at other primary pages). Original summary cites only the generic Apache-2.0 text, which cannot support product-specific claims on hardware, weights or availability. Audit result: Vista's SAMPLING.md suggests GPUs with a minimum of 32 GB VRAM for sampling (with a --low_vram option for GPUs under 80 GB); Cosmos-Predict2.5's multiview inference guide requires at least 8 GPUs with 80 GB each; Vista's README states its repository content is under Apache-2.0 and the Hugging Face model record tags apache-2.0. The 'not found' claims for GAIA, DriveDreamer weights and 1XWM, and the Vista-versus-SVD licence tension, were not re-verified. Replace the single Apache URL with these pages when citing. Version/date: fetched 2026-10-23.
• articleWhere is the safetensors 100,000,000-byte header cap defined, and does ComfyUI use the same limit? Correction: the cap is in the Rust crate's tensor.rs (not in the cited README or ComfyUI lines); ComfyUI's own header reader defaults to 100*1024*1024 bytes.
  Problem. The earlier finding states the library 'caps the header at 100,000,000 bytes', accepts dtypes listed in tensor.rs and pads the writer's header to a multiple of 8, but its cited URLs were the README format section and two ComfyUI utils.py ranges, none of which contain those details. Method. safetensors commit e246a256, tensor.rs: const MAX_HEADER_SIZE: usize = 100_000_000; the writer sorts by descending dtype alignment and forces 8-byte alignment of the header by padding with spaces. ComfyUI commit 46aad33e comfy/utils.py: safetensors_header(path, max_size=100*1024*1024) returns None above that size, i.e. 104,857,600 bytes, a different limit from the library's. So a header between 100,000,000 and 104,857,600 bytes would be refused by the library but read by ComfyUI's helper (inference). Changed since. ComfyUI main now also contains create_safetensors in comfy/utils.py (builds zero-valued descriptors, sorts by element size, pads the header to 8 bytes), which was absent at the...
• articleWhere does Mochi 1's Apache-2.0 licence actually appear? Correction: the cited README lines do not state it; the LICENSE file (Apache-2.0 text) and the Hugging Face card (license: apache-2.0) do.
  Problem. The earlier finding says Mochi's Apache 2.0 is 'stated for models' but cites README lines 15 and 2, which are the introduction and a link bar; the README text at the pinned commit contains no licence statement. Method. Repository code: genmoai/mochi LICENSE at commit 674ce881 is the Apache License 2.0 text. Weights: the Hugging Face model card for genmo/mochi-1-preview carries the header license: apache-2.0 (read from the main branch on 2026-10-23, revision not recorded). Impact. The conclusion (Apache-2.0 for code and the published weights) stands; only the citation was wrong. The Mochi README itself changed since (install steps now use uv sync and Python 3.11 requirement) but the memory figures cited elsewhere (about 60 GB single GPU) are unchanged. Limitations. Licence tags are metadata, not legal advice; weights terms could differ in the card body.
• articleWhat GPU memory do the open image-to-3D releases really need, and does it stop at 24 GB?
  Correction (overstated range). C3-B080-W16-01 says hardware floors stated by the vendors range from about 6 GB to 24 GB. The pinned Hunyuan3D-2.1 README states 10 GB for shape generation, 21 GB for texture generation and 29 GB for shape plus texture in total; TRELLIS.2 states at least 24 GB. So the upper end of the stated range is 29 GB for Hunyuan3D-2.1's full pipeline, not 24 GB. Hunyuan3D-2 (earlier) states 6 GB for shape and 16 GB total. Version/date. Pinned commits fetched 2026-10-23.
• articleC3-B511-02 relayed a creator's Nov 2025 remark that Weavy "just got acquired by Figma". Figma's own site now lists "Figma Weave" in its newsroom navigation as "AI workflows for imagery, video, audio, and more" with a product domain weave.figma.com (R1), and a solutions page states "Every major...
  Corrected 2026-10-23 - see C3-B613-04; the original statements below are kept for history and are superseded where the correction says so. C3-B511-02 relayed a creator's Nov 2025 remark that Weavy "just got acquired by Figma". Figma's own site now lists "Figma Weave" in its newsroom navigation as "AI workflows for imagery, video, audio, and more" with a product domain weave.figma.com (R1), and a solutions page states "Every major AI model, in one place" with image, video, audio and 3D models from Google, Kling, Runway, Luma and others, selectable per workflow step "right from your canvas" (R2). Secondary reports located by search (R3, not read in full) place the announcement around 1 Nov 2025, describe Weave as the new product line built from Weavy, ~20 staff joining, and price estimates of $150-200M (Globes, market sources) - unverified. For agents: name the product "Figma Weave (formerly Weavy)" in current answers; Weavy-era tutorials (model import from fal/Civitai/Replicate, comp...
• articleConversational creative agents inside tools (Artlist agent, Flova skills, Lovart brand kits, Combos.fun game agent): what they automate and where humans still steer
  Creator-reported: statements attributed to the creator(s) come from discovery's reading of auto-caption transcript excerpts; the curator checked the video metadata (title, date, length) and any cited documentation pages, not the spoken claims. Nothing was tested; promotional or sponsored framing applies where stated. Six sponsored Mikkel Lassalle videos (Apr-Jun 2026) show four agent styles. Artlist AI agent: describe a goal in plain language, it picks a model (e.g. suggests 4K-capable models when asked), generates, and proposes next-step edits (moonlight, deeper shadow contrast, anamorphic flare) that you accept or override; same chat then animates a 5 s clip. Flova: "skills" (script-driven, story-driven, reference video, remix, music video, product commercial) wrap a pipeline; his manual path was Nano Banana Pro frame with "use the reference as the absolute identity source" -> GPT Image 2 character sheet -> Seedance 2.0 with both references, 15 s. Lovart: one prompt yields a brand...
• articleAgentic assistants driving Blender, DaVinci Resolve and After Effects for AI film pre-production and editing (GPT-6 Astra, Sept 2026)
  Creator-reported: statements attributed to the creator(s) come from discovery's reading of auto-caption transcript excerpts; the curator checked the video metadata (title, date, length) and any cited documentation pages, not the spoken claims. Nothing was tested; promotional or sponsored framing applies where stated. Dan Kieft (Higgsfield-sponsored, 2026-09-23) on OpenAI's GPT-6 Astra (launched early Sept 2026 per secondary news listings R4/R5, emphasising computer use; primary OpenAI page not read). Uses: (1) ChatGPT + Blender MCP builds a 3D set (a bar) from a prompt and references; he places cameras for shot angles, renders previs frames, has GPT Image 2.5 make them photoreal, then uses them as Seedance 2.5 references - more control than describing shots in text; Astra also built a web viewer of the set so non-Blender teammates can orbit/walk it. (2) Recreating rooms from photos to plan a studio layout. (3) Thumbnail ideation with own portrait, channel examples and inspiration, a...
• articleDecart: priced, API-accessible real-time models - Oasis 3 Preview (text-prompted driving world model) and Lucy 2.5 live video editing at about $0.02 per second
  Decart's API platform (docs fetched 2026-10-23) is one of few places where a world model is a paid, documented API rather than a demo. Oasis 3 Preview (released June 2026, R3): set a scene with a text prompt, then send chunks of exactly four [throttle, steering] actions in [-1, 1] and receive four generated frames for each of three camera streams (left_forward, front, right_forward) over a stateful gRPC session; a new prompt resets the world context; Python SDK decart-oasis; a browser playable preview exists; the docs pitch it as a learned driving simulator for RL, with a PPO Colab (R1). It is driving-specific (vehicle controls, three car-mounted views), not a general walkable world. Pricing (R2): pay-as-you-go, no subscription; Oasis 3 Preview $0.02 per second (a 60 s session = $1.20); Lucy 2.5 realtime video editing (add/replace/remove objects, swap characters, change backgrounds live) $0.02/s at 720p or $0.04/s in fast mode, $0.04/s for batch video; Lucy Restyle 2 $0.01/s; image ...
• articleComfyUI is now agent-operable: Comfy MCP (hosted cloud and open local servers) lets Claude/ChatGPT/coding agents run workflows, and the in-app Comfy Agent builds and edits graphs on the canvas, billed in Comfy Credits
  Comfy MCP (R1): connects external AI agents to ComfyUI over MCP to generate image/video/audio/3D, search models/nodes/templates and run real workflows. Two connections: a hosted Comfy Cloud server at cloud.comfy.org/mcp (OAuth; recommended for claude.ai, ChatGPT and Claude Desktop, added as a custom connector) and a local, fully open-source connection for people already running ComfyUI or working in coding agents (Claude Code via the comfy-cloud plugin from the Comfy-Org/comfy-skills marketplace, Cursor, Codex); both can run side by side with separate sign-ins. The docs advise Mac users wanting open-weight models such as MiniMax H3 or LTX-2.3 to use the cloud connection because those models are too large for practical speed on Apple GPUs. Comfy Agent (R2): in-app agent, generally available on Comfy Cloud (Desktop/local "coming soon", waitlist): finds templates/models/nodes, builds from a template, edits part of a graph while preserving the rest ("Keep the model and sampler settings"...
• articleFrom Mirelo AI GmbH terms (R1): Free Plan users may use the services and outputs only for non-commercial purposes, must visibly attribute Mirelo as sound provider, and outputs may carry a watermark (s2, s7.5); Paid Plan users may use them commercially within the scope of the chosen plan (s2,...
  From Mirelo AI GmbH terms (R1): Free Plan users may use the services and outputs only for non-commercial purposes, must visibly attribute Mirelo as sound provider, and outputs may carry a watermark (s2, s7.5); Paid Plan users may use them commercially within the scope of the chosen plan (s2, s5.2). s7.3: any IP that arises in outputs vests exclusively in the user, but "certain Output may not qualify for intellectual property protection" and use remains subject to plan limits and the Acceptable Use Policy, including restrictions on using outputs to train AI models. s7.4 Training License: the user grants Mirelo a worldwide, royalty-free, sublicensable licence to use Input and Output to train its models and improve services. Paid users can opt out for a personal account via a Studio setting; that setting "does not control organization API keys" - organisations must email legal to confirm or change the arrangement "before submitting content"; opt-out stops future additions and "new thir...
• articleOpen interactive world models you can run or train: Matrix-Game 3.5 (720p, long-horizon memory, first/third person, 40 GB GPU) and ForgeWM (fully open recipe, real-time keyboard/mouse control, trained on 8 GPUs)
  Two open releases extend the open world-model landscape (C3-B399-01). Matrix-Game 3.5 (Riemann Dynamics; HF weights apache-2.0, July 2026; repo Sep 2026) (R1, R2): memory-augmented interactive world model for 720p long-horizon camera-controllable video in first-person and third-person modes. "Patch Memory" lifts past observations into a 3D memory with geometric alignment and visibility-aware retrieval so revisited places stay consistent; static scene and dynamic subjects are separated, with third-person mode taking 0-4 protagonist reference crops for identity. Built on Wan2.2-TI2V-5B with Depth-Anything-3; 5B base (bidirectional) checkpoints plus a three-step distilled first-person causal checkpoint for real-time streaming. Needs one NVIDIA GPU with >= 40 GB VRAM (704x1280 peaks around 40 GB), Linux, >= 64 GB RAM. ForgeWM (arXiv Aug 2026) (R3, R4): an end-to-end open recipe (bidirectional SFT -> teacher-forced causal AR -> consistency distillation -> on-policy DMD) producing 1/2/4-s...
• articleOpen-weight alternatives for upscaling AI video (SeedVR2, FlashVSR+) run in ComfyUI on a 24 GB GPU; generative upscalers trade fidelity for perceived sharpness, so check faces/text after upscaling
  Self-hosted counterpart to commercial finishing tools (C3-B479-01). SeedVR2 (ByteDance Seed): 3B and 7B weights on Hugging Face under apache-2.0; Comfy-Org hosts a repackaged set in fp16, fp8, int8, mxfp8 and nvfp4 precisions (652k downloads) and the community numz ComfyUI wrapper has ~2.9k stars (R1, R5). FlashVSR+ (13 Oct 2026, R2): hybrid local/sparse-global attention gives up to 2x DiT speed and 45% lower peak memory, "enabling spatially untiled 4K VSR on a single 24GB RTX 4090", with reward-weighted distillation for user-preferred one-step output; code, models and training data released (code Apache-2.0; ComfyUI wrappers exist, one GPL-3.0) (R2, R4). A zero-shot pipeline from pretrained parts reaches ~32 fps at 720p on one RTX 4090 but explicitly trades "reconstruction fidelity for strong no-reference perceptual quality" and struggles with hard motion (R3). Practical: (1) a 24 GB card can now upscale generated clips to 4K locally without tiling seams; (2) generative VSR invents...
• articlePost-shoot AI controls on fal H3 Max (Oct 2026): keyframed orbit camera from one still (azimuth/elevation/distance on normalised time), relight from a lighting-sphere image, and recast people from one photo each - with limits and per-second prices
  Camera controls (R3): first frame image + camera_trajectory of keyframes {time 0-1 normalised, azimuth degrees, elevation -90..90, distance in normalised scene units}; first pose held before its time, last held to the end; signed full turns preserved up to 32 turns total azimuth; duration 0.92-15 s (output may run ~0.7 s long); 480P/768P native, 1080P is latent refinement from 768P; the default prompt freezes the scene so "only the camera moves" - i.e. a turntable/orbit tool for set or product coverage from a single still. Price (R6): promotional $0.0325/$0.052/$0.104 per s (480/768/1080p), 35% off until 31 Oct, then $0.05/$0.08/$0.16. Relight (R4): source video up to 15 s plus a sphere render showing the target lighting; video is converted to constant 24 fps and trimmed to the nearest 17n+5 frame count (minimum 56 frames, ~2.33 s), so plan for possible tail trimming in the edit; 480P/768P native, 1080P and 2K via latent upscale; $0.05-$0.32 per s. Recast (R5): source 5-30 s with no...
• articleWhy it matters: many earlier candidates pin prices and schemas on fal (e.g. C3-B501..C3-B507), and vendor pages there carry a "Commercial use" badge, but the governing rights are layered. From fal's terms (R1): (1) Inputs - the customer keeps ownership; fal receives a non-exclusive,...
  Why it matters: many earlier candidates pin prices and schemas on fal (e.g. C3-B501..C3-B507), and vendor pages there carry a "Commercial use" badge, but the governing rights are layered. From fal's terms (R1): (1) Inputs - the customer keeps ownership; fal receives a non-exclusive, non-sublicensable licence to process Customer Input "to provide the Services" (no training licence stated for inputs). (2) Usage Data - defined as anonymised or aggregated data "which may include data based on or derived from Customer Input", which fal owns and may use to "design, develop, and offer Company products, services, and AI models" - so the no-training reading of inputs is not absolute for derived aggregate data. (3) Outputs - no clause assigning output ownership was found; fal "does not represent, warrant, or covenant that any Output Content will be original, will not infringe rights of any third party", outputs may be the same or similar for other users, and use is at the customer's own risk;...
• articleArtlist MCP connector: searching a licensed music/SFX library and generating media from inside Claude or ChatGPT
  Creator-reported: statements attributed to the creator(s) come from discovery's reading of auto-caption transcript excerpts; the curator checked the video metadata (title, date, length) and any cited documentation pages, not the spoken claims. Nothing was tested; promotional or sponsored framing applies where stated. Mikkel Lassalle (Artlist affiliate, 2026-10-18) shows Artlist's MCP integration: from Artlist's MCP page copy the server URL, in Claude add a custom connector (Settings > Connectors > Add custom connector), log in to Artlist, review tool permissions; then ask in natural language for tracks by mood/genre/BPM/duration or sound effects, play previews inline and open them on Artlist; generation tools (e.g. GPT Image 2 image) ask clarifying questions (aspect ratio, quality) and require per-step approval ("allow once") before spending credits. Practical value: licence-cleared music search without leaving the assistant; approval gating prevents silent credit burn. Unverified: ...
• articleHow complete is the Croissant that Hugging Face auto-generates for licence and provenance?
  Corrected 2026-10-23 - see C3-B644-04; the original statements below are kept for history and are superseded where the correction says so. For the four datasets checked (a sample, not a rate), Hugging Face's auto-generated Croissant carries a single dataset-level licence URL copied from the card's licence tag (choosealicense.com form), with no citation, provenance or responsible-AI fields filled, no per-file licences, and a GitHub-issue URL in place of each file's sha256. A dataset without a card licence tag (common_voice_17_0, now distributed via Mozilla Data Collective) gets plain schema.org JSON with no licence or distribution. Auto Croissant is therefore not a provenance or rights record.
• articleHow do I use Real-ESRGAN for video (frames, anime video model, face_enhance, tile), and what are the risks of face enhancement?
  Problem. A cheap, local upscale for video frames. Conditions. Real-ESRGAN (trained on pure synthetic data), BSD-3-Clause code. Method. (1) General video: the repo provides inference_realesrgan_video.py; the anime-video doc example is -i inputs/video/onepiece_demo.mp4 -n realesr-animevideov3 -s 2 --suffix outx2 (realesr-animevideov3 is x4 but can be used for x1-x3; multi-process via --num_process_per_gpu). (2) ncnn route: extract frames with ffmpeg, run the executable on frames, merge back. (3) --face_enhance runs GFPGAN on faces (default off; needs facexlib and gfpgan); -t sets tile size (0 = auto) to save VRAM. Failure modes (inference): the ncnn workflow processes frames independently, so temporal flicker can appear on textured regions; GFPGAN face restoration replaces facial detail and can alter identity, so do not use it on faces that must match an approved reference; the anime video model is for animation, not photoreal faces. Limitations. README does not discuss temporal consi...
• articleWhat does SeedVR2 document for video restoration (one-step diffusion, GPU needs, scripts) and where does it say it fails or over-sharpens?
  Problem. Restoring or upscaling AI-generated or degraded video without temporal artefacts. Conditions. SeedVR / SeedVR2 (ByteDance Seed; ICLR 2026), Apache 2.0; 3B/7B models. Method. SeedVR2 is a one-step diffusion video restoration model with adaptive window attention that adjusts window size to the output resolution. Inference: torchrun projects/inference_seedvr2_3b.py with --video_path, --output_dir, --seed, --res_h, --res_w and --sp_size (sequence parallel GPUs). Stated capacity: one H100-80G handles a 100-frame 720x1280 video; four H100-80G support 1080p and 2K with sp_size=4; tiling tricks were not yet supported. A community ComfyUI node exists (linked in README). Failure modes (README Limitations). Prototype models; not robust to heavy degradation or very large motions; may fail to remove degradation or generate unpleasing detail; on lightly degraded inputs such as 720p AIGC video it may over-generate detail and oversharpen, especially at small resolutions like 480p. Method (...
• articleHow do I outpaint a still with FLUX Fill or SDXL inpainting in diffusers (strength, mask blur, resolution) and which licences apply?
  Problem. Extending a frame beyond its edges while keeping the original pixels. Conditions. diffusers FluxFillPipeline or an SDXL inpainting checkpoint. Method. (1) FLUX.1 Fill [dev]: the docs say the pipeline supports inpainting and outpainting and does not take a strength input; pass prompt, image, mask_image, height, width and max_sequence_length=512; the BFL doc says the mask must be a black-and-white image the same size as the conditioning image. For outpainting (inference): paste the source on a larger canvas, make the mask white over the new area, and set height/width to the canvas size. (2) SDXL Inpainting 0.1: guidance_scale 8.0, 15-30 steps, and strength=0.99 - the card says to keep strength below 1.0; checkpoint is openrail++. (3) General inpaint tuning from the diffusers guide: strength is how much noise is added to the base image, higher gives more difference from the original; blur_factor on the mask processor softens the seam (example 33); padding_mask_crop crops aroun...
• articleWhat do the InsightFace repositories say about the licence of buffalo_l, antelopev2 and other pretrained models versus the MIT-licensed library, and who must be contacted for commercial use?
  Problem. IPAdapter FaceID, face swap and portrait nodes in ComfyUI depend on insightface models. Conditions. deepinsight/insightface commit c6c3af7b (2026-10-19). Cross-checks an earlier round's licence item without repeating it. Method. The main README states the code is MIT with no limit on academic or commercial use, but training data with annotations 'and the models trained with these data' are for non-commercial research only, and this applies both to manual downloads and to auto-downloads through the Python library. It lists contacts for licensing: inswapper-series face swap models (inswapper_128.onnx etc.), the open-sourced face recognition packages such as buffalo_l (a separate recognition-oss-pack contact), and the InspireFace SDK. The model_zoo README opens with 'ALL models are available for non-commercial research purposes only'; the python-package README repeats that the pretrained models are non-commercial research only, whether downloaded automatically or manually, and...
• articleFor a ComfyUI video/image stack, how do code, weights and outputs line up across Wan, Hunyuan, FramePack, LTX, AnimateDiff, RIFE and InsightFace, and which limits are territory, scale, revenue or competing-model clauses?
  Problem. One-glance triage before shipping a workflow or hosted service. Conditions. Synthesis of the sibling findings (pinned revisions in each); not a legal opinion; each cell restates what the cited text says. Method (matrix: code / weights / outputs / hard limits / duties). Wan 2.1/2.2/Animate: Apache-2.0 / Apache-2.0 / no claim by licensor, user accountable / none found / Apache notice duties. HunyuanVideo: licence terms in Tencent Hunyuan Community License / same / Outputs may not be used to improve other AI models, none claimed over content ownership beyond that / excluded territory EU, UK, South Korea; 100M MAU gate; Hosted Service counts / give recipients the agreement, Notice file, change notices, downstream use restrictions. FramePack: Apache-2.0 code / unspecified in sources / unspecified / possible Hunyuan inheritance (unresolved) / none stated. LTX-2 and LTX-2.x: community licences / same / outputs: LTX-Video 0.X says no rights claimed; 2.x text defines Output and bars...
• articleWhen does dynamic batching help or hurt a diffusion/video model server, and which Triton settings trade latency for throughput?
  Problem. Deciding whether to batch concurrent image/video requests on one GPU or run them one at a time. Conditions. NVIDIA Triton Inference Server docs/user_guide/batcher.md at commit d13379e3 (2026-10-22). Method. The docs say the dynamic batcher combines requests into batches to typically increase throughput and should be used for stateless models; it is configured per model (dynamic_batching { }). Tuning loop from the docs: measure latency and throughput with Performance Analyzer under the default config; if latency is within budget, raise the maximum batch size and/or set max_queue_delay_microseconds to a non-zero value, increasing it until the latency budget is exceeded. The delay makes the batcher wait for a full (or preferred) batch but never longer than the configured delay for any request. Preferred batch sizes are discouraged for most models and only worthwhile if a size is much faster. Queue policy lets you cap queue size and set timeouts with reject or delay actions, an...
• articleWhat does the Seed-VC paper claim about zero-shot voice conversion (timbre leakage, similarity, intelligibility), and how big was the evaluation?
  Problem. Voice conversion must keep what was said and how, while replacing who says it; content features often leak source timbre, and removing leakage can raise word error rate. Conditions. Zero-shot: source and target speakers unseen in training; reference audio of a few seconds. Method. Seed-VC uses a diffusion transformer with in-context conditioning on reference audio and an external timbre shifter during training so content features do not carry the source timbre. Measured result (author-reported). On 100 LibriTTS test-clean source utterances and 8 target-speaker references (seed-tts-eval), Seed-VC reached speaker similarity (SECS) 0.8676 and WER 11.99% against ground-truth WER 8.02% and OpenVoice SECS 0.7547; the paper also reports lower CER than baselines. Failure modes. The authors note OpenVoice scored slightly higher on some DNSMOS quality scores, interpreted as a trade-off with timbre fidelity. Limits. Small eval (8 targets), English-centric, objective metrics only, no l...
• articleFrom TikTok Business Help (R1-R4): Symphony Creative Studio is a free AI video tool for TikTok-optimised organic and paid content: text/image/reference-to-video generation, image generation, voiceover avatars (choose from a library filtered by industry, gender, background, region; avatars...
  Corrected 2026-10-23 - see C3-B613-05 and C3-B622-05; the original statements below are kept for history and are superseded where the correction says so. From TikTok Business Help (R1-R4): Symphony Creative Studio is a free AI video tool for TikTok-optimised organic and paid content: text/image/reference-to-video generation, image generation, voiceover avatars (choose from a library filtered by industry, gender, background, region; avatars labelled "Real-human or AI-generated"), product avatars (upload a product or app-screen image and get clips of an avatar holding or modelling it; clothing accessories such as shoes, hats, sunglasses are not supported), translate-and-dub with optional lip sync and captions, an editor, and "Symphony Agent" that turns a prompt into a storyboard and full video. Dreamina Seedance (R2): ByteDance video models integrated into Symphony in variants - Standard (default; broadest features incl. Agent, reference/image/text-to-video and Avatar), Fast (speed; A...
• articleWhat are Seed-VC's practical inputs, latency, singing mode and consent-relevant switches, and under what licence is it released?
  Problem. Choosing an open voice converter for a pipeline requires knowing input limits, real-time latency, singing handling and licence. Conditions. Seed-VC main branch README. Method. Reference speech of 1 to 30 seconds clones a voice with no training; optional fine-tuning needs as little as one utterance per speaker (1 to 30 s files, clean, no background music or noise). Real-time mode has stated algorithm delay about 300 ms plus about 100 ms device delay. Singing: set f0-condition to True and use 30 to 50 diffusion steps; auto-f0-adjust is normally off for singing; semi-tone-shift transposes. A v2 option convert-style uses an autoregressive model to convert accent and emotion (false gives timbre-only conversion); anonymization-only ignores the reference and maps speech to an average voice. Failure modes. Noisy or music-laden source or training audio is explicitly undesirable. Limits. Repo README claims, not independently run. Rights. The repo LICENSE is GNU GPL v3 (copyleft; I re...
• articleFrom the FFmpeg filters documentation (R1): multi-input filters such as overlay share "framesync" options, settable only by name. eof_action - what to do when the secondary input hits EOF: "repeat" the last frame (default), "endall" end both streams, or "pass" the main input through untouched....
  From the FFmpeg filters documentation (R1): multi-input filters such as overlay share "framesync" options, settable only by name. eof_action - what to do when the secondary input hits EOF: "repeat" the last frame (default), "endall" end both streams, or "pass" the main input through untouched. shortest=1 forces output to end when the shortest input ends (default 0). repeatlast=1 (default) extends the last frame of secondary streams until the primary stream ends; 0 disables. ts_sync_mode - "default" takes the secondary frame with the nearest lower-or-equal timestamp to each primary frame; "nearest" takes the absolutely nearest timestamp. Overlay note: frames are taken from each input in timestamp order, so if the inputs start at different timestamps, pass both through setpts=PTS-STARTPTS so they begin at zero. Overlay expressions can use t (seconds) and n (frame number) only when eval=frame (default); with eval=init they evaluate to NAN. Practical recipes (derived): logo or watermark...

More