Local models and model tooling

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

• articleHow do I run a ComfyUI diffusion model with fp8 weights, what does that setting change, and is NF4 (bitsandbytes) still a supported route?
  Problem. Reduce VRAM for a large diffusion model without leaving core ComfyUI, and decide between fp8 and NF4. Conditions. ComfyUI commit 9fe7ef76 (comfyui_version.py shows 0.42.0). Method. Two core routes exist. (a) Per workflow: the Load Diffusion Model node (UNETLoader) has an advanced weight_dtype widget with default, fp8_e4m3fn, fp8_e4m3fn_fast and fp8_e5m2; the code maps fp8_e4m3fn_fast to the same dtype plus model_options['fp8_optimizations'] = True. (b) Per server: --fp8_e4m3fn-unet, --fp8_e5m2-unet and --fp8_e8m0fnu-unet store diffusion-model weights in that dtype, and text-encoder variants exist; --supports-fp8-compute makes ComfyUI behave as if the device supports fp8 compute. Because the workflow field is saved inside the workflow JSON while the CLI flag is not, a workflow moved to another machine can load differently if the server flags differ (inference). NF4. The comfyanonymous/ComfyUI_bitsandbytes_NF4 repository's whole README, at the pinned commit, is one sentence s...
• articleAnswers "should colour scopes/transforms apply to EXR depth or data channels?" from primary docs. OpenEXR (R1): channels are named strings (e.g. R, Z, yVelocity); only R, G, B (intensities) and A (opacity; colour channels premultiplied by convention) have predefined interpretations. Data types:...
  Answers "should colour scopes/transforms apply to EXR depth or data channels?" from primary docs. OpenEXR (R1): channels are named strings (e.g. R, Z, yVelocity); only R, G, B (intensities) and A (opacity; colour channels premultiplied by convention) have predefined interpretations. Data types: HALF for regular image data, FLOAT "where the range or precision of 16-bit number is not sufficient (for example, depth channels)", UINT for "discrete per-pixel data such as object identifiers". Compression: lossy DWAA/DWAB quantise DCT components of RGB data; B44/B44A are lossy for HALF; PXR24 keeps HALF and UINT exact but rounds FLOAT to 24 bits (~3e-5 relative error) and "works well for depth buffers"; lossless ZIP/PIZ-type codecs avoid any change - so put data passes (Z, normals, IDs, motion) in lossless or PXR24-compressed parts, not DWA. Data window vs display window: pixel data may extend beyond (overscan) or within the display window. Colour identification: OpenEXR does no colour tran...
• articleCan a ComfyUI diffusion model be exported to ONNX and TensorRT for speed, what is portable between machines, and which features break?
  Problem. Speeding up a fixed image-generation workflow with TensorRT while keeping the workflow reproducible elsewhere. Conditions. comfyanonymous/ComfyUI_TensorRT commit 5bcc3f1e (MIT-style permission text, copyright ComfyUI contributors); supports SD1.5, SD2.1, SD3, SDXL, SDXL Turbo, SVD, SVD-XT, AuraFlow and Flux. Method. The conversion node exports the loaded model with torch.onnx.export (opset_version=17, dynamic axes) to a temporary model.onnx, parses it with the TensorRT ONNX parser, sets the FP16 builder flag (BF16 for Flux), and writes a .engine file; a timing_cache.trt next to the code speeds later builds. Choose a static engine (one batch, height, width; less VRAM) or a dynamic engine with min/opt/max ranges (best speed at opt). Engine file names encode the shape (stat-b-..., dyn-b-min-max-opt-...). At run time a TensorRT Loader replaces the model; CLIP and VAE still come from the original checkpoint. Portability. The README says the engine is specific to your GPU, so the...
• articleWhat does ComfyUI-GGUF let me do with quantised diffusion models (Flux, SD3.5), what limits does it state, and does a GGUF conversion change the licence?
  Problem. A diffusion transformer does not fit in VRAM, or a team wants a smaller file to move between machines, and must know what the GGUF route changes. Conditions. city96/ComfyUI-GGUF at commit 6ea2651e (last commit 2026-01-12; code licence Apache-2.0 text). Targets transformer/DiT models (the README names Flux and SD3.5 Large quants) and a quantised T5 text encoder. Method. (1) Put .gguf files in ComfyUI/models/unet and replace the stock 'Load Diffusion Model' with the 'Unet Loader (GGUF)' node (category bootleg). (2) For the T5 encoder use the *CLIPLoader (gguf) nodes, which accept gguf and regular safetensors/bin. (3) Install the gguf Python package. (4) Keep ComfyUI recent enough for custom ops. (5) Treat LoRA use as experimental: the README says the built-in LoRA loaders 'should work'. What it changes. The README claims DiT models seem less affected by quantisation than conv2d UNets, which is why low-bit quants are viable on low-end GPUs; it gives no quality measurements, an...
• articleWhich AI outputs does Adobe indemnify? Firefly models vs partner models (Google, OpenAI) in Firefly Creative Production for Enterprise
  Primary help page (enterprise scope only). Partner models offered in Firefly Creative Production for Enterprise include Google Gemini 3 (Nano Banana Pro), Gemini 3.1 (Nano Banana 2), Veo, and OpenAI GPT image models incl. GPT Image 1.5 and GPT Image 2; Firefly's web docs also list video generation with Veo, Luma, Runway and Kling partner models (navigation). Indemnity: for ELIGIBLE, GENERALLY AVAILABLE Google and OpenAI media models surfaced there, Adobe indemnifies certain COPYRIGHT infringement claims under the customer's Adobe terms; this differs from Firefly-model indemnification and does NOT cover trademark, publicity or privacy claims; it can be void if the customer knew or should have known an output infringed, lacked input rights, or ignored safety features; beta/preview/early-access models are excluded and coverage "should not be assumed" for other models, providers or Adobe products. Third-party models wired in via custom nodes/actions with your own API key fall under the ...
• articleHow HeyGen gates real-person digital twins: consent video levels, identity matching, and the enterprise "skip consent with indemnity" option
  Closes the HeyGen consent follow-up. Every video-based "digital twin" requires proof the depicted person agreed; photo-based and prompt-generated avatars (no identifiable real person) have no consent step (consent_status null). Three access levels (developer docs): (1) webcam recording on HeyGen's hosted consent page - all customers; the API (POST /v3/avatars/{group_id}/consent) returns a link valid 24 h for the subject, who reads a statement (with a spoken code in the interactive flow); (2) uploading a pre-recorded consent video - enterprise only; (3) skipping consent entirely - enterprise only with a signed indemnity agreement. Requirements: the same person as the training footage (identity-matched), one clearly visible face, the provided statement read (validated semantically, any language), under 35 s per developer docs (help centre says under 30 s - version discrepancy), MP4/MOV/WebM, good light and audio, no recording of a screen; if creating a twin of someone else, that perso...
• articleHyperFrames: an open-source "HTML in, MP4 out" renderer built for coding agents - how creators use it for motion graphics and what it requires
  Primary README (heygen-com/hyperframes, Apache-2.0, Node >= 22; releases daily - v0.8.188 on 2026-10-23). It renders HTML, CSS, media and seekable animations into deterministic MP4 via a CLI, a Studio/playground, or as a rendering core, and is designed to be driven by coding agents: install as a Claude Code plugin (marketplace add heygen-com/hyperframes; /hyperframes) or via "npx skills add"/"npx hyperframes skills update" for Codex, Cursor, Gemini CLI, Copilot and others. Seven skills; the core set is /hyperframes (render contract and references for film, captions, overlays, motion, look), /hyperframes-cli (scaffold, check, inspect frames, edit timeline, preview, render locally or in the cloud, publish), /hyperframes-registry (~450 ready-made editable blocks: transitions, titles, kinetic type, captions, charts, maps, 3D, camera moves) and /media-use (music, SFX, images, voiceover, avatars, generated video, transcripts, cutouts, grades - via HeyGen services or free local tools); opt...
• articleFrom Higgsfield API documentation: Lifecycle (R4) - every model uses the same authenticated asynchronous flow; a POST returns status "queued", a request_id, a status_url and a cancel_url; poll or receive a webhook; input media can be uploaded to Higgsfield storage first. Concurrency (R2) - the...
  From Higgsfield API documentation: Lifecycle (R4) - every model uses the same authenticated asynchronous flow; a POST returns status "queued", a request_id, a status_url and a cancel_url; poll or receive a webhook; input media can be uploaded to Higgsfield storage first. Concurrency (R2) - the primary limit is the number of requests processing at once; above it, generation requests "are still accepted" and wait in the queue, while some non-generation endpoints (e.g. custom Soul ID reference training) return 429 with Retry-After; successful responses carry no remaining-capacity headers, so clients must track request_ids, cap submissions with a worker pool/semaphore, separate polling from submission, and back off with jitter. (Concurrency tiers by recent top-ups: 2/10/20/40 - C3-B516-01.) Billing (R1) - only successful generations are charged; requests ending "failed" or "nsfw" are refunded automatically if credits were reserved; a queued request can be cancelled (and refunded) only b...
• articleMidjourney in Aug-Oct 2026: V8.2 edit model (instruction edits, 4 image references, in/outpainting), "Thinking Mode" reruns, shared folders
  Primary update feed. 27 Aug 2026: first V8.2 image EDIT model opened to everyone - edit images by instruction, generate from up to 4 image references at once ("replacing omni-reference"), inpainting and outpainting, works with personalization, moodboards and style refs (which may need extra prompt direction); usable via prompt-bar attachments, lightbox Edit, the Edit tab, or --edit on Discord; image quality update on 28-29 Aug. Sep 2026 (alpha site): style live-previews using fast models, default parameters per user/folder, improved in/outpainting that changes only selected pixels so repeated edits do not degrade the image, Korean UI. 7-8 Oct 2026: "Thinking Mode" test - rerun a V8.2 job "with thinking": Midjourney inspects the image for misses against the prompt (wrong objects, layout, anatomy, text) and regenerates; reported to help prompt accuracy, typography and coherence; shared folders with collaborators. Not found in the feed: a new Midjourney video model release in this wind...
• articleSeedance 2.5 limits from BytePlus's own API reference: 30 images / 10 videos / 10 audio, 4-30 s, 1080p (10-bit HEVC), draft tasks, and no direct real-face references
  Primary ModelArk docs (model id example dreamina-seedance-2-5-260628) settle creator-reported numbers. Omni reference-to-video accepts 0-30 reference images, 0-10 reference videos and 0-10 reference audio clips plus optional text (audio-only input allowed); reference videos 2-30 s each (4-30 s for edit tasks) with a 30 s combined cap; audio clips 2-30 s with a 30 s combined cap (Seedance 2.0: max 3 videos/3 audio, 15 s caps; first/last-frame and multi-image modes: 2.5 takes 1-30 images vs 1-9 for 2.0). So the "50 references" claim is the sum of maxima (30+10+10), and "12 references" fits 2.0 (9 images + 3 videos/audio). Output: Seedance 2.5 default 720p, also 480p and 1080p (1080p output uses 10-bit colour and H.265/HEVC); Seedance 2.0 supports up to 4K; 2.0 fast/mini 720p max. A separate LAS doc says 1080p is unsupported for 2.5 in reference-image scenarios (R2) - check per endpoint. Duration: 2.5 default -1 (auto) or 4-30 s (2.0: 4-15 s); edit tasks require duration -1, ratio adap...
• articleCredit-saving direction habits (Sep 2026): motion-only video prompts, don't name things you don't want rendered, and let an LLM "director" write second-by-second blocking
  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. Youri van Hofwegen. Higgsfield production (R1, sponsored): wasted credits are "almost never the model's fault"; generate the character sheet and environment once (GPT Image 2 for photorealism), then drop BOTH in as references for each shot; when a shot already has a start frame, the video prompt "only has to cover motion and camera"; use lens/angle language deliberately (e.g. floor-level, steep up-angle, 24 mm wide to make a subject look powerful); and avoid naming objects you do not want - "anything you name as an object is something the model will try to render" (naming "the camera" can put a camera in frame). Settings he used: Seedance 2.5 at 10 s, 1080p, 16:9; also ...
• articleCoverage note with small facts: 11 more sponsored or self-promotional Mikkel Lassalle uploads (Feb-Mar 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. Concrete items: Artlist platform tour (2026-03-12, R6) - claims 50M creators; video tools with negative prompts, multiple references, multi-output, four cinematic styles and clip chaining into a long scene; custom voice cloning from about 10 seconds of audio (consent rules apply). Artlist "Big Ad" challenge with a $60,000 prize (R12) - contests as a sponsor marketing tactic. Enter.pro (R2): an "AI development agent" test building a deployable app with payments in about an hour; it recommended Gemini 3 Pro or Claude Sonnet 4.5 for the app's text generation; he configured a freemium plan (5 free shot plans/month, unlimited on Pro). Framia Pro (R1): agent platform with cre...
• articleByteDance OmniHuman v2 vs v1.5 on fal (Oct 2026): single image + audio to talking/performing video; v2 takes 4-300 s audio with an optional multilingual prompt at $0.105/s, v1.5 caps audio at 30 s (1080p) / 60 s (720p), adds a speaker mask and turbo mode at $0.16/s
  Hosted on fal (R4): three OmniHuman endpoints active - original, v1.5 (updated 21 Apr 2026) and v2 (updated 15 Oct 2026). All take one image of a human figure plus an audio file and return a video whose length equals the audio; billing is per second of audio/video ("duration ... as used for billing"). v2 schema (R1): required image_url + audio_url; audio wav/mp3, max 15 MB, 4-300 s; image jpeg/png/webp/bmp/tiff/gif/heic/heif, max 30 MB, aspect ratio 0.4-2.5, 300-6000 px per side; optional prompt for scene, movement and camera in Chinese, English, Japanese, Korean, Mexican Spanish or Indonesian; no resolution, mask or turbo parameters exposed. v1.5 schema (R2): audio must be under 30 s for 1080p or under 60 s for 720p (the schema says 720p is "faster and higher in quality"), turbo_mode trades slight quality for speed, and mask_url restricts speaking to the person in the white mask area - the only multi-person control in this family on fal. Rendered price lines (R3, 23 Oct 2026): v2 $...
• articleComfyUI security (Jul 2026 advisories): four high-severity flaws fixed in v0.28.0 - why an exposed ComfyUI must be patched and kept off the open network
  Four GitHub security advisories published 2026-07-15 for Comfy-Org/ComfyUI, all affecting versions before 0.28.0 and fixed in 0.28.0: two stored cross-site-scripting issues (CVSS 8.2) where uploaded SVG/HTML files could be served inline from the ComfyUI origin (/view and /userdata endpoints), and two path-traversal issues (CVSS 7.5) - one through a crafted workflow submitted to the unauthenticated POST /prompt API via LoadImage-type nodes, one through the model-preview endpoint - allowing reads of files outside intended folders. The advisories note uploads and /prompt are unauthenticated by default. Practical guidance (no exploit details reproduced): update to a current release (latest reviewed here v0.42.1, C3-B396-01); do not expose ComfyUI on public interfaces (avoid --listen 0.0.0.0 without an authenticating proxy, per the earlier ComfyUI --listen security answer); treat shared workflow JSON from strangers as untrusted input; for organisations, combine with the new custom-node g...
• articleTool snapshots from mid-2026 sponsored walkthroughs: Meshy texturing options, Lumina canvas planning, Artlist preset apps, Pippit "story series" pitch
  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. Meshy (2026-07-30): image-to-3D "quick generate" or agent mode; standard vs "smart topology"; texture step with options to remove baked lighting, generate PBR maps, and texture resolution 2K/4K/8K. Lumina canvas (2026-08-04): describe the story -> auto storyboard of individual shots -> "director's desk" to set framing, composition, camera movement and key shots before generating (claims no queue waiting). Artlist "AI apps" (2026-07-01): 70+ one-click presets in categories (viral, editing tools, camera movement, product, typography, social, collabs) - e.g. product photo -> lifestyle shot or packaging mockup. Pippit Story Studio with Seedance 2.5 (2026-08-20): script/scre...
• articleAsset-library production tools (Artlist Studio, Apr 2026): characters with generated angles and voices, multi-angle locations, shot/angle/lighting presets and structured prompts
  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, 2026-04-21). Artlist Studio project model: create a character from one reference image, then "generate angles" for multi-view consistency and pick a voice (custom voices announced as coming); a library of pre-made characters; locations generated from a prompt plus multiple angles for location consistency; in prompts you @-tag characters and locations. Prompt helpers: angle presets, shot presets (close-up, extreme close-up, full, wide, extreme wide), lighting presets, and a "structured prompting" mode with fields for characters, location, action, composition, style and mood. Generate a style frame, "set start frame", then in a directin...
• articleThe common shape of "director mode" AI film agents (CapCut Director Mode, OiiOii, Lumina canvas, SekoTV): plan, assets, storyboard, gated clip generation
  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. Four sponsored Mikkel Lassalle walkthroughs (May-Aug 2026) show the same agent pipeline: one-line idea (+ optional style reference image) -> agent writes a plan/script and asks format questions (short vs long, aspect, language, an "emotion keyword") -> role-named sub-agents generate character sheets, environments and a storyboard (OiiOii: art director, character designer, scene designer, storyboard artist) -> human approval gate -> clip generation -> final approval -> export to an editor. CapCut Director Mode example: the agent generated four 4-second Seedance 2.0 clips at 720p for a music-video sequence and offered alternative audio tracks before the final "approve del...
• articleAgentic all-in-one AI video platforms (SekoTV) and image-to-3D platforms (Hi3D): what the workflows look like
  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. SekoTV (sponsored, 2026-09-22): a "creation agent" with skill templates (cinematic one-take short, character concept board, TVC product ad, master director, etc.) takes a one-line idea, proposes a configuration (tone, aspect) and three plot options, writes asset prompts, generates assets with GPT Image 2, waits for "confirm visual assets", then writes and runs the final one-take video prompt with sound; models offered include Seedream 5.0 Pro, GPT Image 2.5, Nano Banana Pro. Human checkpoints shown are confirmation gates rather than prompt edits. Hi3D (sponsored, 2026-09-29): text-to-image (Nano Banana 2 recommended; a clean grey background helps) then image-to-3D with ...
• articleHow are creators building their own AI filmmaking tools with coding agents and MCP connectors, and what are the practical limits?
  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 (2026-08-01) describes "vibe-coding" three tools with a coding agent (Claude Code; Fable 5 or cheaper Opus 5): define the problem (his AI-film workflow had too much back-and-forth between an LLM and a generation platform), answer three questions (what, why, which features), have an LLM turn them into a master build prompt ("meta prompting"), revise it heavily, then iterate in the coding agent. Results: (1) a story/character/location "world builder" web app where users connect their own Claude API key and their Higgsfield account via MCP (so the site runs at no cost to him), which writes a structured script PDF, generates character sheets (GPT Image 2) and loca...
• articleIs Fish Audio S2 "open-source", and what can it do for AI voice-over (emotion tags, cloning)?
  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 (2026-03-10, Fish Audio link) calls Fish S2 "probably the strongest AI open-source model" and demonstrates bracketed emotion/breath tags (whispering, soft or loud voice, inhale, exhale), cloning "from as little as 15 seconds of audio", multilingual output, streaming and an API; workflow: draft a trailer script, polish with ChatGPT, generate, iterate quickly, cut picture to the narration. Licence check on Hugging Face (2026-10-23): fishaudio/s2-pro and the newer fishaudio/drama-3 (2026-10-17) are under a "fish-audio-research-license", and s1-mini is CC-BY-NC-SA-4.0 (gated); the fish-speech code repo shows NOASSERTION. So the weights are openly downloadabl...
• articleHow can creators cut AI video generation cost: cheaper models by task, draft-then-upscale, and API vs subscription?
  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 (2026-10-13, Higgsfield) gives three tactics with Higgsfield credit prices on that date: (1) pick the cheapest model that gets 80-90% of the result for the task: a matched shot cost 12 credits on Kling 3.0 vs 72 on Seedance 2.5; for anime he preferred MiniMax H3 (30 credits, 37.5 "max") over Seedance 2.5 (180); his free "model selector" (an LLM over his own test notes) recommends models per prompt, duration and resolution. (2) Seedance 2.5 draft mode: iterate at 480p (45-60 credits per try in his numbers) and pay once to render the chosen take at 1080p (180-240 credits), e.g. 3 iterations cost about 420 credits instead of 720. (3) API vs subscription: pay-per-...
• articleHow do you build a consistent multi-character cast with LoRAs (dataset, captions, training steps) for image and video models, per a July 2026 local workflow?
  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. Mickmumpitz (2026-07-17) gives five LoRA rules: (1) a made-up trigger word absorbs everything not captioned; (2) vary what should generalize (include wide shots, lighting, poses, clothing), or the character only exists as trained (e.g. close-ups); (3) decide what stays fixed versus editable; (4) captions decide editability: captioned clothes stay changeable but must then be re-prompted; (5) captions are reverse prompts, so match the target model's prompt style (long natural language for Krea 2; structured JSON with bounding boxes for Ideogram 4). Pipeline: an edit-model "character creator" generates a varied synthetic dataset from one or two images (add close-ups of det...
• articleHow does a local generative video upscaler (Wan 2.2 tiled) compare with Topaz and FlashVSR, and what settings matter?
  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. Mickmumpitz (2025-11-25): Wan 2.2 low-noise model in 81-frame batches with overlap blending and tiling; FP8 for 24 GB+ GPUs, otherwise GGUF sized to VRAM (Q5 for 12 GB; he uses Q8); LightX2V low-noise LoRA for fewer steps; optional realism LoRA; an ESRGAN-style pre-upscaler; auto-captioned prompt. Settings: target 2K as a good default; a quality pass adds ~30% quality for ~2x time (skip at 4K); "creativity" (denoise) low for fidelity, higher for degraded footage (risk visible tiles/subject change); save frames as fallback because stitching very long 4K outputs can crash ComfyUI. ~10 minutes for 81 frames; RunPod RTX 5090 ~$0.89/h. His comparisons: his 2K beat Topaz 4K o...
• articleWhich metrics, baselines and benchmark protocols do the 4D generation papers use, and which reported numbers can be compared across papers without being misled?
  Partially verified: the curator re-read none of the cited sources; S1, S3, S8, S13, S11, S5, S6, S15, S16 rest on the producer's pinned reading. The papers use different metric families and protocols: CLIP R-Precision plus human votes (MAV3D), CLIP Score plus pairwise user study (4D-fy), LPIPS/CLIP/FVD on a seven-object synthetic benchmark (Consistent4D, DreamGaussian4D, L4GM, SV4D), CLIP-F/CLIP-O and image metrics on a 20-case held-out set (Diffusion4D), and view/frame FVD variants plus FV4D (SV4D family). Cross-paper tables carry copied or inconsistent rows: the DreamGaussian4D row in L4GM and SV4D tables (0.16/0.87) equals the Consistent4D row and differs from DreamGaussian4D own Table 3 (0.12/0.92); Consistent4D FVD is 1133.44 in two papers and 1133.93 in a third; L4GM calls the benchmark eight animations while Consistent4D and its uid file list seven. User-study percentages point in opposite directions in MAV3D and 4D-fy. The one independent-protocol re-run found (SV4D 2.0 on i...
• articleCan I get an animated file I can open in Blender, Unreal or Unity (mesh sequence, PLY sequence, glTF, Alembic) out of these text/image/video-to-4D systems, and what did the authors actually release?
  Partially verified: the curator re-read none of the cited sources; S1, S3, S2, S12, S11, S6, S7, S5, S17, S18, S9, S4, S14 rest on the producer's pinned reading. Among the eight systems, only DreamGaussian4D documents a command-level path to a DCC tool: its README exports a per-frame OBJ mesh sequence and loads it in Blender with the Stop-motion-OBJ add-on. MAV3D describes a mesh-sequence conversion in the paper but has no available code. L4GM ships weights and code whose inference script writes only mp4 renders; a PLY writer exists in its repository but is not called, so a per-frame PLY sequence is an untested route. SV4D / SV4D 2.0 releases sample view-by-frame image grids, and no fitting or export code is visible by path name. No source read mentions Alembic, glTF, FBX or USD output.
• articleWhat do text-to-4D, image-to-4D and video-to-4D generation mean in practice, and for MAV3D, 4D-fy, DreamGaussian4D, Consistent4D, SV4D/SV4D 2.0, L4GM and Diffusion4D, what goes in, what representation comes out, and is it something I can run?
  Partially verified: the curator re-read none of the cited sources; S1, S3, S11, S13, S15, S5, S8, S4, S6, S12, S17, S9 rest on the producer's pinned reading. "4D" here means a 3D object or scene that changes over time and can be rendered from new viewpoints; every system read produces a generated asset, not a captured one. Three routes appear in the primary sources: (a) per-asset optimisation against diffusion priors, by score distillation (MAV3D, 4D-fy, Consistent4D) or, for DreamGaussian4D, by fitting Gaussian deformation to a generated driving video with image-diffusion guidance, (b) a diffusion model first generates a grid of views by frames, then a 4D representation is fitted to it (SV4D, SV4D 2.0, Diffusion4D), and (c) a feed-forward network that reconstructs per-frame 3D Gaussians from a video (L4GM). Output representations are dynamic NeRFs or Gaussians; only the DreamGaussian4D abstract mentions animated meshes (per-frame, candidate 02) and no abstract describes a rigged as...
• articleWhat frame counts, resolution rules and frame rates do the READMEs of LTX-Video, CogVideoX, Mochi, Wan and HunyuanVideo require, so a shot of a given duration can be specified correctly?
  Partially verified: the curator re-read none of the cited sources; S12, S17, S6, S8, S19, S4, S3 rest on the producer's pinned reading. Frame counts have model-specific forms: Wan 4n+1 (defaults 81 and 121); LTX-Video 8n+1 with sizes divisible by 32 (padding otherwise); CogVideoX 8N+1 with N<=6 (default 49) for the original models and 16N+1 with N<=10 (default 81) for CogVideoX1.5; HunyuanVideo 129 in every listed setting; Mochi: README example uses 480x848 at 31 frames, rule not stated in the README. Frame rates stated: Wan 16 fps (24 for TI2V-5B), CogVideoX 8 fps (1.5: 16 fps), Wan2.2 TI2V-5B 24 fps. Duration = frames / fps, e.g. CogVideoX 49/8 = about 6 s as the README says.
• articleWhat do the Wan2.1 and Wan2.2 READMEs and configs state about memory, offloading flags, supported sizes, frame counts and frame rate for each variant?
  Partially verified: the curator re-read none of the cited sources; S1, S2, S6, S3, S4, S5 rest on the producer's pinned reading. Wan2.1: T2V-1.3B supports 480P (832*480); 14B supports 480P and 720P; OOM is handled with --offload_model True and --t5_cpu; the time/peak-memory table is an image. Wan2.2 (commit 1ea34ff48f87168174e12956e200b1d908b1c5ff): A14B variants state at least 80GB VRAM for the 1280*720 command, TI2V-5B states at least 24GB at 1280*704 with offload, dtype conversion and T5 on CPU; defaults are 81 frames at 16 fps (A14B family) and 121 frames at 24 fps (TI2V-5B). The Wan2.2 SUPPORTED_SIZES list in code does not include sizes beyond those in the README. The efficiency table with per-GPU seconds and peak GB is a PNG, so the actual numbers are not in text.
• articleBeyond the 28 decorators in server.py, which HTTP routes do the user, settings, model, custom-node, subgraph, node-replacement, internal and asset modules register at the pinned commit, and which only read versus change server state?
  Partially verified: the curator re-read none of the cited sources; S1, S2, S3, S4, S5, S8, S10, S12 rest on the producer's pinned reading. At commit 1a82e6d, files read: 44 further route registrations sit outside server.py's own 28: 20 from UserManager (incl. AppSettings), ModelFileManager, CustomNodeManager, SubgraphManager and NodeReplaceManager (each also served under /api, so 40 URLs), 5 under /internal, 16 asset routes and 3 export/task routes with literal /api paths. By HTTP verb 27 are reads (GET/HEAD) and 17 change state (users.json, comfy.settings.json, userdata files, asset database rows, asset scans, terminal-log subscription). The verb is not a clean read/write marker: some GETs create directories or write access times, so G43's read/write split needs these corrections.
• articleWhat is ComfyUI's V3 node API (schema, entry point, versions), how does the server load it next to V1 nodes, and which stability and secrecy limits should an agent check before writing or installing a V3 node?
  Partially verified: the curator re-read none of the cited sources; S17, S23, S22, S21, S18, S19, S20 rest on the producer's pinned reading. V3 nodes subclass io.ComfyNode, declare an io.Schema in define_schema and implement a classmethod execute; a module exposes them through an async-or-sync comfy_entrypoint returning a ComfyExtension whose async get_node_list lists the classes. The loader tries NODE_CLASS_MAPPINGS first, then comfy_entrypoint, registering nodes by schema.node_id and skipping a module with a warning if the entry point misbehaves. The versioned API is young: latest and v0_0_2 are STABLE False while a v0_0_1 adapter is STABLE True, which conflicts with the guide's statement that v0_0_2 is the first version. Hidden inputs can hand a node the Comfy account API key, so installing a V3 or V1 node is installing code that may see credentials.
• articleAnswers "how do procedural systems keep feet in contact and the body supported?" from engine documentation. Plant (contact) settings (R1): speed_threshold - "Bone is considered planted below this speed" (speed from an FK speed curve or graph root motion); unalignment_speed_threshold - speed at...
  Answers "how do procedural systems keep feet in contact and the body supported?" from engine documentation. Plant (contact) settings (R1): speed_threshold - "Bone is considered planted below this speed" (speed from an FK speed curve or graph root motion); unalignment_speed_threshold - speed at which the foot becomes fully unplanted, the range between the two approximating the foot's roll phase (the docs themselves note this "feels innaccurate most of the time"); distance_to_ground - within this distance of the plant plane the bone is planted and fully aligned; unplant_radius and unplant_angle - linear and angular deviation that release the lock; replant_radius_ratio/replant_angle_ratio - below these fractions of the unplant limits the foot replants (hysteresis); min/max_extension_ratio bound how far the leg chain may extend; separating_distance keeps the feet apart; ankle_twist_reduction limits slope alignment twist; adjust_heel_before_planting; reconstruct_world_plant_from_velocity...
• articleONNX Runtime v1.31.0 exposes a per-session deterministic-compute switch (config key session.use_deterministic_compute / OrtApi::SetDeterministicCompute), while the CUDA execution provider defaults to exhaustive cuDNN convolution algorithm search and TF32 on - both outside that...
  ONNX Runtime v1.31.0 exposes a per-session deterministic-compute switch (config key session.use_deterministic_compute / OrtApi::SetDeterministicCompute), while the CUDA execution provider defaults to exhaustive cuDNN convolution algorithm search and TF32 on - both outside that switch.
• articlePyTorch v2.14.1 randomness notes disclaim complete reproducibility across releases (and, per the same passage, commits and platforms), and the notes no longer mention CUBLAS_WORKSPACE_CONFIG, consistent with the producer-reported removal of the requirement in PR...
  PyTorch v2.14.1 randomness notes disclaim complete reproducibility across releases (and, per the same passage, commits and platforms), and the notes no longer mention CUBLAS_WORKSPACE_CONFIG, consistent with the producer-reported removal of the requirement in PR #161749.
• articleWhich ComfyUI command-line flags control output caching, what exactly does each do, and what is the default?
  At the pinned commit there are four mutually exclusive cache options plus --high-ram: --cache-ram [GB [GB]] (RAM-pressure caching, stated to be the default), --cache-classic, --cache-lru N, --cache-none, and --high-ram, which sets classic caching. Only --cache-none stops cross-run reuse of outputs; the others all use the same input-signature keys and differ in what they evict. The default was not always RAM pressure: at two earlier commits the default was classic. The commit/release that switched the default was not located.
• articleHow does group offloading work in diffusers v0.41.0 for video models (block_level vs leaf_level, streams, disk), which combinations raise errors, and which numbers exist for it with hardware?
  Group offloading (ModelMixin.enable_group_offload, DiffusionPipeline.enable_group_offload, hooks.apply_group_offloading) moves groups of layers between CPU/disk and the accelerator. The source at the pin turns several documented cautions into hard errors: block_level needs num_blocks_per_group; record_stream needs use_stream; use_stream needs a non-CPU backend with streams; Accelerate hooks on the same module conflict; model/sequential CPU offload on a pipeline that has any group-offloaded component raises; disk offload with torchao tensors raises. Documented softer cautions: weight-dependent device casting may clash, streams need about 2x CPU memory of the model, block_level with streams is forced to 1 block per group, VAEs with streams and tiling need a dummy forward pass. Timing/memory numbers linked by the guide are a Flux table (no GPU named) and a Wan table (no GPU, size or resolution named), both from a proof-of-concept PR; no video-model benchmark with named hardware was fou...
• articleAddendum to C3-B396-01: what ComfyUI "governance enforcement for custom nodes" does
  PR #16167 (merged 2026-10-06): normal installs are unchanged (GOVERNANCE_REQUIRED False; initialize returns early). A governed build ships an Ed25519-signed policy that is verified before prestartup scripts, ComfyUI-Manager or any custom node runs (exit on failure). Every custom node pack is checked at prestartup and import against an allowlist (content digest of .py/.pyd/.so/.dll/.dylib/.js etc.) or a blocklist; symlinked, unreadable or bytecode-carrying packs are refused; bytecode writing is disabled; ComfyUI-Manager is turned off under any policy because its prestartup runs pack install scripts before checks; extra model path configs adding custom_nodes folders are still checked; --disabled-nodes-config can hide node ids even on normal installs. Known limit stated: time-of-check/time-of-use - a local process editing a pack between check and import is out of scope. Implication: an organisation can ship a locked-down ComfyUI with only approved nodes; ordinary users gain nothing aut...
• articleDawDreamer docs at e0daff3c state both that automation granularity is the block size and that automation data is interpreted at audio rate; the render loop calls automateParameters once per block with the buffer size, consistent with block-wise application (interpolation code not...
  DawDreamer docs at e0daff3c state both that automation granularity is the block size and that automation data is interpreted at audio rate; the render loop calls automateParameters once per block with the buffer size, consistent with block-wise application (interpolation code not read).
• articleWhat memory thresholds does the RAM-pressure default use, how did they change between the flip commit and v0.23.0, and where do the PR description and code disagree?
  At the flip commit 5aa5ccc9 the default thresholds were active 25% of RAM (4-32 GB) and inactive 75% (12-96 GB). Three days later commit b30e980a (PR #14089, listed in the v0.23.0 notes) lowered them to active 10% (2-10 GB) and inactive 100% (max 96 GB), with the stated reason that many users have 'it just fits' workflows. The PR #13802 description says the second argument defaults to 25%, which contradicts the 75% in the merged code. The same release also reverts the 'disable non-dynamic smart memory' step of #13802 (PR #14183). Net behaviour at v0.23.0 was not read from main.py at the tag.
• articleWhat steps, learning rates, class-image counts and VRAM figures do the DreamBooth and textual-inversion papers and the diffusers v0.41.0 docs state, and where do they differ?
  As stated: DreamBooth paper ~1000 iterations, prior-loss weight lambda=1, learning rate 1e-5 (Imagen) or 5e-6 (Stable Diffusion), 3-5 images, ~1000 class samples (fewer possible), about 5 minutes on one A100 for Stable Diffusion. diffusers examples/dreambooth README: 5e-6 with 400 steps (no prior), 5e-6 with 800 steps and 200 class images (prior-preservation), 2e-6 for the 12 GB variant, LoRA DreamBooth 1e-4 and 500 steps, with the text encoder needing at least 24 GB. Textual inversion paper: batch size 4, base lr 0.005 scaled to 0.04, 5,000 steps on 2 V100s, about two hours per concept; diffusers README: 5.0e-04 with --scale_lr and 3000 steps. Hugging Face's own blog experiments report 400 steps at 2e-6 for objects and ~1200 steps at 1e-6 for faces. The values differ by source and model, and are used as stated; none is a universal setting.
• articleWhere do licence statements, dependency notices, dataset provenance and weight file formats disagree or leave gaps for these 3D generators?
  Four items need curator attention: (1) TripoSG advertises MIT but its NOTICE lists code derived from Tencent HunyuanDiT (Tencent Hunyuan Community License notice) and the BRIA RMBG-1.4 background-remover whose card says non-commercial; the README downloads RMBG weights. (2) TRELLIS-500K metadata says MIT while the DATASET.md describes assets curated from Objaverse(XL), ABO, 3D-FUTURE, HSSD and Toys4k and states the dataset is CSV metadata; asset-level licences were not read. (3) Hunyuan3D repos ship .ckpt (pickle-capable) next to safetensors; 2.1 ships .ckpt and .bin; TripoSR ships model.ckpt; the licence texts do not address file formats. (4) The Tencent Hunyuan 2.0 HF card links its licence at LICENSE.txt while the HF file list shows LICENSE.
• articleFor the audio-capable world models and sync-focused sound tools found here, what do the repositories and model cards actually say about licence and weights per component, and where do they disagree?
  HelixWorld (Noiz AI) is the one audio-visual world model with a published checkpoint: code and the HF card say Apache-2.0, but the repository NOTICE says the LTX-2 code and model assets inside it are governed by the LTX-2 Community License (paid licence for entities with at least US$10M annual revenue; fine-tuned weights count as derivatives), the paper says HelixWorld builds on LTX-2.3, and the Gemma text encoder has its own terms. WorldSonus is CC BY-NC 4.0 with .pt weights (pickle-capable). V-AURA code is MIT but its checkpoint is a tar.gz with no stated weight terms. Kling-Foley publishes a paper and a CC-BY-4.0-labelled benchmark dataset but no code or weights were found.
• articleWhich of the scene-graph systems have released code, under what stated licence, and what third-party models, keys or data does each need?
  At the observed commits, Hydra (BSD-style LICENSE), ConceptGraphs (MIT) have permissive code licences, while the 3DSceneGraph repository carries a non-commercial research-use software licence and its data sits behind a separate click-through agreement that was not accessed. ConceptGraphs code is MIT but running it needs third-party segmentation and vision-language checkpoints plus an OpenAI API key, whose terms are separate. A permissive code licence does not state anything about weights, data or outputs.
• articleWhen a model's GitHub LICENSE, Hugging Face licence metadata and model-card prose disagree, does Hugging Face or the vendor say which governs, and what does the HF licence field officially mean?
  At the pinned revisions I found no Hugging Face page and no vendor text (Whisper, VibeVoice, AudioSeal) that ranks the GitHub LICENSE, the HF front-matter licence and the card prose against each other. HF's docs describe the licence field as the owner's way to 'let other users know' the permissions, shown on the model page and used for filtering; custom terms belong in a LICENSE file in the repo. HF's Terms of Service say content carrying a licence notice is intended to stay under that licence and that downloaded content is subject to the Terms 'and/or' the terms accompanying it. So the safe agent reading is: record each channel separately with its revision, call the HF field 'metadata says', and do not pick a winner.
• articleDo the Cosmos repositories, model cards and docs restate the NVIDIA Open Model License accurately, and do the dated versions that appear in model-card front matter differ from the current text?
  Three dated agreement texts are readable without accepting anything: the current page (October 24, 2025), the text embedded in the Cosmos-Predict2.5-14B card (September 23, 2025, an abridged rewording), and the text embedded in the Cosmos-Tokenize1 card (January 6, 2025). The January text uses 'Model Derivatives', has no Special-Purpose Model clause, no 'Guardrail' definition and no Separate Components section, and its guardrail sentence has no 'substantially similar Guardrail' exception. GitHub READMEs and the docs licence page restate the licence only by name and link. One Cosmos 3 model repository ships a LICENSE file that matches the October 2025 text. Model-card 'Important Note' paraphrases of the guardrail clause differ between cards.
• articleBefore I use, host, fine-tune or redistribute a Cosmos model that is under the NVIDIA Open Model License, what does each clause of the current agreement text say, and which of my questions does it not answer?
  At the dated text 'Last Modified / Version Release Date: October 24, 2025' the agreement has 11 numbered sections. It grants a perpetual, worldwide, no-charge, revocable licence to use, modify, make derivatives of, sell and distribute the Model; allows redistribution with a copy of the agreement and a Notice line; requires 'Built on NVIDIA Cosmos' for Cosmos Models in several uses (including training another AI model with them or their outputs); states NVIDIA claims no ownership of outputs; and ends your rights automatically on litigation or on bypassing a Guardrail without a substantially similar one. It has no clause on hosting as a service by that name, no express patent licence (the word 'patent' appears only in the litigation-termination sentence) and no operative commercial-use clause beyond the preamble statement. This is a reading of text, not legal advice.
• articleThe agreement ends my rights if I bypass a Guardrail without a substantially similar one. What do the agreement text, the Cosmos docs and the Cosmos code at a pin say about whether the guardrails can be turned off?
  Three source kinds disagree on tone, not necessarily on fact: the agreement (October 2025) ends rights if you bypass a Guardrail without a substantially similar one; the Cosmos docs state the Cosmos-1.0-Guardrail model 'cannot be disabled' in the diffusion and autoregressive pipelines; and cosmos-predict1 at one commit defines a --disable_guardrail command-line argument and forwards it to the pipeline. The docs sentence and the code are about overlapping but not proven-identical pipelines; the code path was not traced or run. 'Substantially similar Guardrail appropriate for your use case' is not defined in the text.
• articleDid the AudioSeal vendor ever publish a separate weights licence file, what do the GitHub and Hugging Face histories show, and how can an agent check release-scoped weights terms?
  Yes, in an earlier revision. At git 785b3b11 the AudioSeal repo has a LICENSE_weights file carrying Creative Commons Attribution-NonCommercial 4.0 text and a README that says code is MIT and weights CC-BY-NC 4.0. Commit 29565154 ('Re-update of license (#28)', committer date 2024-04-02) removes LICENSE_weights and edits the README; at HEAD e63a8a0e the only LICENSE is MIT and the weights are covered only by a README news line, while the README License section still names only code. The Hugging Face repo has no licence file in its tree, and its README gained license: mit in a later commit than its first two README versions.
• articleCan the Meta open speech-translation toolchain (SeamlessM4T v2, SeamlessStreaming, SeamlessExpressive, NLLB-200) be used in a commercial dubbing pipeline, and which licence applies to which component?
  Seamless code (and the W2v-BERT 2.0 encoder, UnitY2 aligner, toxicity tools and the text-only part of mExpresso) is MIT; SeamlessM4T v1/v2 and SeamlessStreaming weights are CC BY-NC 4.0 (the HF v2 card agrees); Seamless and SeamlessExpressive weights fall under a separate Seamless Licensing Agreement limited to noncommercial research that also bars commercial use of outputs, terminates on IP litigation against Meta and requires indemnity; NLLB-200 models are CC BY-NC 4.0 while fairseq code is MIT. Commercial dubbing with these weights is not permitted by these texts.
• articleWhere does a Hugging Face model repository tagged creativeml-openrail-m keep its licence text, and how do I check what it contains?
  The CompVis stable-diffusion-v1-4 card (133a221b) declares creativeml-openrail-m, allows redistribution and commercial use only with the same use restrictions passed on, and points to a Hugging Face Space for the full text, with no LICENSE-named file at the repository root; under the same metadata value one derived repo ships the CreativeML Open RAIL-M text (dated 22 Aug 2022) as LICENSE.md, while pt-sk/stable-diffusion-1.5 ships an unfilled MIT licence template as license.txt.
• articleWhat do the model cards of four open TTS releases say about voice consent, impersonation, provenance marking and licence, and where do card statements and licence metadata disagree?
  At the pinned revisions, the VibeVoice-1.5B card declares MIT in its front matter while also limiting the model to research use and listing voice impersonation without explicit, recorded consent as out of scope, and claims an audible disclaimer and watermark in outputs (vendor statements); the XTTS-v2 card (Coqui Public Model License) advertises cloning from a 6-second clip with no consent text; the F5-TTS card declares CC BY-NC 4.0 with no use-policy text.

More