Local models and model tooling

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

• articleSDXL-Turbo: 1 step, guidance 0, 512x512 and the img2img steps x strength rule
  SDXL-Turbo: 1 step with guidance_scale=0.0 (no CFG, no negative prompt), 512x512 preferred. For img2img keep steps x strength >= 1 (2 steps at 0.5 strength = 1 step). Licence: non-commercial sai-nc-community.
• articleAnimateDiff v1-5-2: baseline diffusers recipe and the clip_sample rule
  AnimateDiff v1-5-2 baseline: SD1.5 finetune plus motion adapter, DDIMScheduler with clip_sample=False, linspace timesteps and steps_offset 1, 16 frames, guidance 7.5, 25 steps. Disable sample clipping in any clipping scheduler.
• articleStable Video Diffusion XT: 25 frames at 576x1024, limits and licence
  SVD-XT turns one 576x1024 image into 25 frames (about 4 s max) with no text control; about 180 s per clip on an A100 80GB. Outputs are watermarked by default. Non-commercial use is under the SVD community licence; commercial use needs Stability AI terms.
• articleLTX-2.3: checkpoints, 32-pixel and 8k+1-frame rules, licence
  LTX-2.3 ships 22B dev and 8-step CFG=1 distilled checkpoints with latent upscalers. Use width and height divisible by 32 and 8k+1 frames (pad with -1 otherwise). Licence: LTX-2 Community License, not open source.
• articleHunyuanVideo 1.5 720p T2V in diffusers: attention backend and default settings
  HunyuanVideo 1.5 720p T2V example: bf16, CPU offload, VAE tiling, 121 frames, 50 steps, 24 fps. Install kernels and use attention_backend("_flash_3_hub") on H100/H800 or "flash_hub" elsewhere; the default backend also works.
• articleAnimateDiff-Lightning: step checkpoints, scheduler settings and ComfyUI setup
  AnimateDiff-Lightning has 1/2/4/8-step motion modules for SD1.5. Use EulerDiscreteScheduler with trailing timesteps and linear betas, guidance 1.0 and steps equal to the checkpoint; the authors like 3 steps on the 2-step model and Motion LoRAs at 0.7 to 0.8.
• articleHugging Face model cards: lineage, licence and version metadata
  Check base_model (and base_model_relation) to tell an original model from a fine-tune, adapter, quantization or merge; license: other means read license_name and license_link; new_version links to a newer model.
• articleAnimateLCM: diffusers recipe for fast SD1.5 AnimateDiff video
  AnimateLCM pairs its MotionAdapter with an SD1.5 model, LCMScheduler (linear beta schedule) and its t2v LoRA at weight 0.8; the card example makes 16 frames at guidance 2.0 in 6 steps.
• articleHow do prompted frontier LLMs compare with fine-tuned small models at next-state prediction in text environments (including TextWorld), and when did predicted rollouts stay consistent with the real environment?
  Corrected 2026-10-24 - see C3-B1131-01; the original statements below are kept for history and are superseded where the correction says so. Li et al. (arXiv 2512.18832 v2) measure exact-match (EM) next-state accuracy in five text environments. Prompted models (zero-shot or 3-shot) reached 0.00-17.70% zero-shot and 11.43-49.12% 3-shot on TextWorld, the narrative-game engine; fine-tuned Qwen2.5-7B and Llama-3.1-8B (trained on 40K GPT-4o trajectories per environment) reached about 70.5%. On ALFWorld/SciWorld fine-tuned models reach 99%+ but zero-shot frontier models 35-65%. Consistency of rollouts (success when replaying world-model actions in the real environment divided by real success) was 0.91-0.96 in the three structured environments for the fine-tuned Qwen world model and 0.67 on WebShop. Fine-tuned models are not "the LLM itself" prompted; the authors say so by contrasting prompting with SFT.
• articleWhich terms govern the text-safety models bundled with or called by Cosmos-Transfer1: meta-llama LlamaGuard-7b, NVIDIA Aegis Defensive 1.0 and Llama-Guard-3-8B?
  Corrected 2026-10-24 - see C3-B1125-02; the original statements below are kept for history and are superseded where the correction says so. The HF Cosmos-Transfer1-7B snapshot at ef4dc8e3 caches meta-llama/LlamaGuard-7b (Llama 2 Community License per Meta's gating text) and two revisions of NVIDIA's Aegis adapter (card: governed by the Llama 2 Community License). The GitHub code at 5005e823 instead calls meta-llama/Llama-Guard-3-8B (Meta gating text: Llama 3.1 Community License; NVIDIA ATTRIBUTIONS.md embeds the Llama 3 text), which is not in the bundle. The upstream Meta licence and use-policy files were not copied into the bundle snapshot, and the Meta READMEs and licence files are gated; the licence text comes from the public gating-prompt metadata. No legal conclusion is drawn.
• articleWhat does 'world model' mean in the 2018 Ha & Schmidhuber paper that many later papers cite, and how does its Vision/Memory/Controller design work?
  Corrected 2026-10-24 - see C3-B1125-01; the original statements below are kept for history and are superseded where the correction says so. In Ha & Schmidhuber (arXiv v4) a world model is a learned generative model of an RL environment: a VAE (V) compresses each frame to a latent z, an MDN-RNN (M) predicts the next z given the action, and a tiny linear controller (C) picks actions from z and M's hidden state. The paper shows C can be trained inside M's generated 'dream' and then transferred, and that this fails when C exploits M's errors unless the model's temperature is raised. The meaning is a latent-dynamics model for control on small game environments, not a video or 3D generator. Schmidhuber's 1989-90 'model network' reports are the cited lineage; the phrase 'world model' is the 2018 paper's usage.
• articleHiggsfield (consumer/creator tiers): you keep Outputs and may use them commercially, but inputs and outputs may train Higgsfield models, outputs may not train any other model, and the underlying model can be swapped without notice
  Corrected 2026-10-24 - see C3-B1096-01; the original statements below are kept for history and are superseded where the correction says so. Primary Higgsfield Terms of Use (last updated 26 Jul 2026). Ownership: Company "does not claim ownership" of Inputs or Outputs and does not restrict commercial use; output rights survive cancellation or account deletion and can be sublicensed to clients (s4.4). Training: Your Content, Inputs and Outputs "may be used by Company to train" its and affiliates' models; opt-out is only by deleting content or the account, going forward; Enterprise Agreement customers are excluded from training and treated as confidential (s4.4). Marketing reuse only for content you made public on the service or with consent. Likeness: uploading a person's name/likeness/voice means you represent you hold all releases and consents, including for generating Outputs (s4.2(b)). Outbound restriction: you may not use the Service, Outputs or derived data to train, fine-tune or...
• articleWhat do ComfyUI-Manager's security_level, install flags and model-download gate actually permit, and where do docs and README disagree?
  Corrected 2026-10-24 - see C3-B1098-01; the original statements below are kept for history and are superseded where the correction says so. Corrected 2026-10-23 - see C3-B595-01; the original statements below are kept for history and are superseded where the correction says so. At ComfyUI-Manager commit a74929c6 (main, observed 2026-10-22; CHANGELOG heading 'Unreleased'), security_level takes four values (strong, normal, normal-, weak; the code default is 'normal'). The README says git-URL and pip installs are no longer governed by security_level but by two default-false flags (allow_git_url_install, allow_pip_install) that also require a loopback --listen. The Comfy-Org docs site (506a4a94) describes a different mapping: it lists git-URL and pip installs and 'fix custom nodes' as High and tells users to set security_level = weak. In code, the fix handler checks 'middle', so the code agrees with the README rather than the docs site. Model downloads through install_model pass three g...
• articleLuma's public API in Oct 2026 still documents only Ray 2 / Ray 2 Flash - while ComfyUI deprecated its Ray 2 nodes and Luma is absent from current AA video boards
  Corrected 2026-10-24 - see C3-B976-01; the original statements below are kept for history and are superseded where the correction says so. Primary API docs list two video models - Ray 2 (ray-2) and Ray 2 Flash (ray-flash-2) - with text-to-video and image-to-video via start/end keyframes, resolutions 540p/720p/1080p/4K, durations such as 5 s, loop and aspect-ratio options, and "concepts" (camera-motion presets) retrievable from /generations/concepts/list; the FAQ says API outputs carry no watermark and may be used commercially. Signals that this surface is ageing: ComfyUI deprecated its Luma Ray 2 partner nodes in v0.39.0 (5 Oct 2026, R3), and Luma models do not appear in the Artificial Analysis T2V/I2V boards captured today (C3-B438-01). We did not find (in this pass) API documentation for a newer Luma video model; Luma's consumer app may differ. Practical: before building on Luma's API, confirm model roadmap and support; older archive guidance on Luma camera concepts/keyframes (Ray...
• articleBefore loading a shared ComfyUI workflow, what should an agent check, and which checks are documented versus only proposed?
  Corrected 2026-10-24 - see C3-B1070-01; the original statements below are kept for history and are superseded where the correction says so. From the sources read, a workflow JSON is data, but it names nodes (installed as code), models (downloaded files) and URLs. Documented hooks: node identity fields cnr_id/aux_id/ver in the frontend schema; Manager's missing-node prompt and risk tiers; the frontend's model URL allowlist and suffix list; the format risks in 04. Hash fields exist but are optional. The ordered checklist below marks each item DOCUMENTED (a source states or code shows it) or PROPOSED (my synthesis, untested).
• articleCan an agent run these action-following benchmarks itself: what code, data and checkpoints does each need, in which file formats, and what do the repositories say about terms?
  Corrected 2026-10-24 - see C3-B1046-02; the original statements below are kept for history and are superseded where the correction says so. Per component at the pinned commits: GameWorldScore (code Apache-2.0 text in its directory, MIT at root; needs five external helper checkpoints in .pth/.model/.weights formats); WorldScore (code MIT, dataset on HF with metadata mit and a WorldScore-Dataset.zip in its file list, evaluation checkpoints .pth/.pt/.pkl fetched by wget, leaderboard submission by results or by videos); MineWorld (code MIT, metrics/ folder, .ckpt checkpoints named in the README tree but taken down per the README and the HF API returned 401 on 2026-10-22); CameraCtrl (Apache-2.0 code, .ckpt weights on HF); Yume (Apache-2.0 code, weights on HF per README, processed dataset listed as not yet provided, Yume-Bench data not located); GameCraft (Hunyuan Community License excluding EU, UK, South Korea; .pt weights; no evaluation code in the root listing). No repository text rea...
• articleIs MiniMax H3 open-weights, and what does its repository say about what is and is not released?
  Corrected 2026-10-24 - see C3-B1044-01; the original statements below are kept for history and are superseded where the correction says so. The MiniMax-AI/MiniMax-H3 GitHub repo (head d21241f0, 2026-08-15) links Hugging Face weights and a "MiniMax H3 Community License" (custom, not OSI; GitHub detects no licence). Its README states the H3-Context-IR preprocessing workflow is not in the open-source release (an API reproduces it), the initial release offers full-attention inference only (sparse attention later), and the H3-Encoder uses the full Qwen3-VL-32B weights (layer-50 hidden states), implying a heavy local footprint. It also ships an "h3-prompt-writing" agent skill (Markdown) usable in Claude Code and other harnesses. This reconciles creator claims: Mikkel Lassalle presented H3 as a hosted, cheaper-than-Seedance model (C3-B335-08); Mickmumpitz called it "free open-source" and used it for video cleanup (C3-B338-02). Both are partly true: open weights under a community licence, p...
• articleWhich open image-to-3D models can studios use commercially, and are there territorial limits (TRELLIS.2, TripoSR, Hunyuan3D)?
  Corrected 2026-10-24 - see C3-B1038-04; the original statements below are kept for history and are superseded where the correction says so. Checked 2026-10-23: Microsoft TRELLIS.2 (GitHub MIT, pushed 2026-07-10; HF TRELLIS.2-4B card mit) and VAST TripoSR (MIT) are permissive. Tencent Hunyuan3D 2.1 (and 2.0, card "tencent-hunyuan-community") community licence states in capitals that it "DOES NOT APPLY IN THE EUROPEAN UNION, UNITED KINGDOM AND SOUTH KOREA"; the licence grant is "for the Territory only"; use, reproduction, distribution or display of the Works, Output or results outside the Territory is unlicensed; and licensees whose products exceeded 1 million monthly active users must request a licence. The same pattern appears in MiniMax H3 (also excluding the US) (C3-B364-01). ComfyUI ships official tutorials for TRELLIS.2, Hunyuan3D-2/2.1, Pixal3D and TripoSplat (docs index captured in C3-B370), so availability in a tool is not a licence grant. For hosted generators (Meshy, Rodin)...
• articleAgent-controlled Blender and DaVinci Resolve via MCP (Oct 2026): what the leading open servers do, and their security, telemetry and licensing catches
  Corrected 2026-10-24 - see C3-B1037-03; the original statements below are kept for history and are superseded where the correction says so. Follows up the creator claims in C3-B393-03. Blender: the dominant server is the community "mcp-for-blender" (renamed; not affiliated with the Blender Foundation; MIT): an MCP server plus a Blender add-on that opens a socket inside Blender to receive and execute commands; one-line installer configures many clients (Claude Desktop/Code, Codex, Cursor, VS Code, etc.). Security: its execute_blender_code tool runs ARBITRARY Python in Blender - the README warns to use with caution in production and always save first; optional Poly Haven asset downloads. Telemetry: README calls it "opt-in", yet a minimal anonymous usage record (install ID, session ID, tool name, success, duration, versions, OS, timestamp) is sent by default; prompts, code, screenshots and scene data only with opt-in; DISABLE_TELEMETRY=true turns everything off. DaVinci Resolve: davinc...
• articleWhat are Gemini Omni Flash's documented capabilities, limits, prompting rules, watermarking and price in the Gemini API?
  Corrected 2026-10-24 - see C3-B978-01; the original statements below are kept for history and are superseded where the correction says so. Gemini API docs (captured 2026-10-23): gemini-omni-1.1-flash generates video with audio from text, image (reference/first frame), first+last frame interpolation and subject reference images, edits videos conversationally (multi-turn via the Interactions API) and extends videos at the tail. Output resolution 360p, 720p (default) or 1080p (upscaled); aspect 16:9 or 9:16. Limits: uploaded videos for editing/extension must be 10 s or less; extension appends to the end only; you cannot extend an uploaded talking clip to add dialogue (except via multi-turn on model-generated videos); voice editing and audio references unsupported; video references max 3 clips of up to 3 s each (audio ignored); no multi-video reasoning; no system instructions, temperature, top_p, stop sequences or negative-prompt field (write "Do not do X" in the prompt); YouTube URLs n...
• articleIs there an open-licence dereverberation tool with a paper reference I can cite, and what does it need?
  Problem: Reduce room reverb on recorded speech with open code. Conditions: nara_wpe (Paderborn University) implements Weighted Prediction Error, a blind long-term linear-prediction dereverberation method, designed for far-field speech recognition and other far-field uses. Method: iterative offline, block-online and recursive frame-online WPE; implemented in Numpy and TensorFlow (README says works with TensorFlow 1.12.0); tested with Python 3.7-3.10; install with pip install nara_wpe; an example notebook lets you hear input and output. Licence: MIT, Copyright 2018 Paderborn University. Cite Drude et al., ITG 2018. Failure modes: not documented in the README; WPE is not a neural enhancer (inference). Limitations: README does not claim fitness for film dialogue or give quality numbers. Version/date: master branch README fetched 2026-10-24.
• articleCan one ComfyUI instance use several GPUs, and do the docs say to run one instance per GPU?
  Problem: scaling renders over several GPUs. Conditions: fb3befd source plus docs.comfy.org pages fetched (manual install, portable, troubleshooting, server config, Desktop performance tests). Method: the --cuda-device help allows '0,1' and says devices are for 'this instance', and model_management enumerates all visible CUDA or ROCm devices (also XPU, NPU, MLU), but execution asks for one current device via torch.cuda.current_device(). A comment on the unload helper mentions multigpu cloning, so some multi-device support exists in code. Documented answer: I found no docs.comfy.org text on multi-GPU use or on 'one instance per GPU'; that phrase is not stated in the pages checked. Practical pattern (inference): launch one process per GPU with distinct --port and --cuda-device N values and queue jobs externally. Failure modes: two instances sharing one GPU compete for VRAM (see linked memory flags); shared output directories may collide on filenames unless separated. Limitations: do no...
• articleWhat do ComfyUI --cuda-device and --default-device do, and what does the Windows single-GPU default mean?
  Problem: choosing which GPU(s) a ComfyUI process sees. Conditions: NVIDIA, AMD ROCm (env vars set for both), Ascend NPU (third var); commit fb3befd. Method: --cuda-device takes a comma-separated id list or 'all'; main.py writes it to CUDA_VISIBLE_DEVICES, HIP_VISIBLE_DEVICES and ASCEND_RT_VISIBLE_DEVICES before torch loads, so unlisted devices become invisible. --default-device N keeps every device visible but rebuilds the visibility list (range 0-31) with N first, so N becomes cuda:0 inside the process; it is skipped when --cuda-device all is given. Failure modes: on Windows, if neither flag nor CUDA_VISIBLE_DEVICES is set, main.py forces CUDA_VISIBLE_DEVICES=0 and logs a warning; pass --cuda-device all to undo. Ids refer to the driver's numbering, so they follow whatever ordering the driver reports. Limitations: reading of code, not run on a multi-GPU machine. Version: fb3befd.
• articleWhat does MagicArticulate release (skeleton model vs skin weights), and what does its dataset cleaning say about skin-weight data quality?
  Problem. Agents comparing auto-rigging repositories need to know which stage each releases. Conditions. Seed3D/MagicArticulate (CVPR 2025 per README), Apache-2.0 code. Method. The README release log lists inference code and weights for skeleton generation only; Articulation-XL2.0 (about 48K models, split 46.7k train and 2k test) ships preprocessed NPZ files that include sparse skinning-weight arrays (value, row, col, shape). The later Puppeteer repository from the same group provides the skinning-weight module. Data-quality lesson. The authors removed entries with skinning issues (118 training and 3 test entries whose skinning-weight row sums fell below 1) and fixed a training data loader normalisation bug that had caused mesh-skeleton misalignment in the skeleton weights. Failure modes. Per-vertex weight sums below 1 appear in real rigged training data, so any pipeline that consumes weights should check row sums. Limitations. README head only was read; the README may state other li...
• articleCan I use Puppeteer (Seed3D, NeurIPS 2025) to predict skin weights for a mesh and skeleton I already have, what does it need, and what licences apply?
  Problem. Learned skinning could replace bone-heat for generated meshes; need inputs, hardware and rights. Conditions. NVIDIA CUDA setup: README pins Python 3.10, PyTorch 2.1.1, CUDA 11.8, flash-attn and pytorch3d. Method. The skinning module takes .obj meshes plus rig .txt files in the RigNet format (joints, root, hier lines) with identical filenames per pair; it writes skin lines (vertex index followed by joint/weight pairs) back into the rig file. The README says inference needs 4.2 GB of VRAM and that complex topology takes longer. Export to FBX uses bpy 4.2.0. Rights. Repository LICENSE is Apache License 2.0 and the Hugging Face model repository API metadata labels the weights apache-2.0; the Articulation-XL2.0 dataset is labelled cc-by-4.0. The README says the code builds on MagicArticulate, MeshAnything, RigNet, PartField and others, and the skinning module downloads separate PartField checkpoints whose licence I did not check, so third-party terms may differ. Evidence on qual...
• articleWhat preprocessing limits and licence apply to Wan-Animate when an actor video drives a character image?
  Problem: the open Wan-Animate model needs pose and face streams extracted from the actor video, and mismatches cause artifacts. Conditions: Wan2.2 repo; animation mode (character mimics actor) or replacement mode (character replaces the person in the video). Method: preprocessing outputs src_pose.mp4 and src_face.mp4, plus src_bg.mp4 and src_mask.mp4 in replacement mode; the paper abstract says body motion is replicated with skeleton signals and expressions with implicit facial features. In animation mode, enable retargeting when body proportions differ, and check the preprocessing results before generation. In replacement mode retargeting is off by default. Failure modes: significant body-proportion mismatch can cause deformation; mask extraction targets single-person videos and can fail with several people; a lower fps setting can cause choppiness. Rights: weights are Apache 2.0, with a use clause barring content that violates laws or harms individuals. Limitations: not run by me....
• articleFor the open LivePortrait method, how should the driving video be prepared, and can I use it commercially?
  Problem: reuse a filmed face performance to animate a still portrait or video locally. Conditions: KwaiVGI/LivePortrait repo, inference.py with -s (source) and -d (driving) arguments. Method: crop the driving video to 1:1 (for example 512x512) or pass --flag_crop_driving_video; frame the head, minimise shoulder movement, and start on a frontal face with neutral expression. Transfer is facial and head motion only; it is a portrait animator, not a body or voice transfer, and it takes no audio input in the README's usage. Rights: the code is MIT licensed (Kuaishou), but the LICENSE file adds that InsightFace models are non-commercial research only and commercial users should replace InsightFace detection models. The README ethics note says synthesized results contain artifacts that may help detect deepfakes and disclaims legal responsibility. Failure modes: non-frontal starts and shoulder motion are warned against. Limitations: I did not run the code. Version/date: README fetched at st...
• articleDoes the cache stay portable across machines and repos: what do symlinks, shared Xet blobs and HF_HUB_DISABLE_SYMLINKS change, and is hf_transfer still used?
  Problem. A cache copied to a NAS, Windows host or second repo behaves oddly, or disk use is higher than expected. Conditions. Recent huggingface_hub with hf_xet. Method. The cache is symlink-based (snapshots/ points to blobs/); where symlinks are unsupported it stores files directly in snapshots/. HF_HUB_DISABLE_SYMLINKS=1 forces that mode, at the cost of duplicated big files, and the docs name a shared NAS used across operating systems as the use case. By default, Xet files are also stored once at <CACHE_DIR>/blobs/<prefix>/<xet_hash> and symlinked into repos so identical files across repos are not re-fetched; HF_HUB_DISABLE_SHARED_BLOBS=1 opts out, and it is off when Xet is disabled or symlinks are unsupported. Fast transfer is hf_xet (installed with huggingface_hub since 0.32.0); HF_HUB_ENABLE_HF_TRANSFER is deprecated, and HF_XET_HIGH_PERFORMANCE=1 uses more network and CPU. Failure modes. Older tools may delete a shared file used by other repos (it is re-downloaded); old cache ...
• articleHow do I run a generation worker with Hugging Face offline mode, and what does it do to a cache hit versus an API call?
  Problem. Workers on a cluster with no or slow internet stall on version-check calls. Conditions. Cache already populated (ideally by a pinned prefetch). Method. Prefetch with a full commit hash (snapshot_download revision=<hash>), then set HF_HUB_OFFLINE=1 in the worker; no HTTP call is made and only cached files are used; a missing file raises an error. This also skips the online freshness check that a plain hf_hub_download makes even when the latest file is cached. Failure modes. Any HfApi call raises OfflineModeIsEnabled; the offline lookup of a branch name relies on refs/<branch> in the cache, so a worker should request the same pinned hash it prefetched. Limitations. Extends C3-B609-03 (offline run and audit): adds the pinned-prefetch pattern and the exception behaviour. Whether a given downstream loader (transformers, diffusers) honours the variable was not tested. Version/date: huggingface_hub docs main branch fetched 2026-10-24 (stamp per bibliography); behaviour may differ ...
• articleA model repo is gated: what is the access-request flow, which token scope should a production worker use, and what does a failed fetch mean?
  Problem. Download of a gated model fails or a worker lacks access. Conditions. Repos with access requests enabled. Method. A human logs into their Hugging Face account, clicks Agree on the repo page (shares username and email, plus any form fields, with authors), and approval is automatic or manual per the authors. Scripts must then authenticate with a user token (hf auth login, login(), or HF_TOKEN). For production, a member requests access, then creates a fine-grained token with read access to that model, so the token cannot read other private repos; fine-grained tokens can also be shared in an organization. Failure modes. A rejected user cannot request again; authors can reset a decision, which revokes access and asks the user to agree again; huggingface_hub raises GatedRepoError (HTTP 403) when access is missing. Org policies may require approval or only fine-grained tokens (403 for read/write tokens). Limitations. Access is per user account and terms may restrict redistribution...
• articleWhich token does huggingface_hub use when several are available, and how do I stop a pipeline sending my token everywhere?
  Problem. A batch job authenticates with an unintended (e.g. write-scope) token or leaks it. Conditions. Scripts, containers, CI. Method. Order: a token passed as token= to a call; HF_TOKEN environment variable, which overrides the stored token; the token stored by hf auth login at $HF_HOME/token (or HF_TOKEN_PATH). Once logged in, all Hub requests, even ones needing no auth, send the token; set HF_HUB_DISABLE_IMPLICIT_TOKEN=1 to stop that. Use one token per app/machine so one can be revoked alone. Failure modes. Hardcoding the token in code or notebooks; reusing a write token for a read-only download. Limitations. Docs state the precedence of env over stored token; the exact rank of an explicit token= argument versus HF_TOKEN is not stated in these excerpts. Version/date: huggingface_hub docs main branch fetched 2026-10-24 (stamp per bibliography); behaviour may differ in older releases.
• articleWhere does the Hugging Face cache live, what is its on-disk layout, and which environment variables move it?
  Problem. Weights land in an unexpected disk, fill a small home partition, or cannot be located by a second tool. Conditions. Python huggingface_hub and tools built on it. Method. Default cache is ~/.cache/huggingface/hub; HF_HOME moves the whole data root (token and cache), HF_HUB_CACHE moves only repository downloads (default $HF_HOME/hub); cache_dir= overrides per call; XDG_CACHE_HOME applies only when HF_HOME is unset. Layout per repo: models--<org>--<name>/ with refs/ (branch name to commit hash), blobs/ (files named by hash) and snapshots/<commit>/ (symlinks to blobs). Read files via the snapshots path, never copy from blobs. Failure modes. Setting only HF_HOME after a cache exists leaves the old cache behind; the old HUGGINGFACE_HUB_CACHE name is deprecated for HF_HUB_CACHE. Limitations. Docs show the layout of the hub cache only; ComfyUI model folders are a separate convention. Version/date: huggingface_hub docs main branch fetched 2026-10-24 (stamp per bibliography); behavio...
• articleHow do I pin a Hugging Face download to an exact commit so a generation pipeline reproduces, and why is a branch or tag name not enough?
  Problem. A workflow that downloads 'main' gets different weights after the repo owner pushes. Conditions. Any hf_hub_download, snapshot_download or hf download call. Method. Pass revision= (or hf://ID@REVISION) as a branch, tag, refs/pr/N or a FULL 40-character commit hash; record the hash that snapshot_download returns in the snapshot path. Short 7-character hashes are not accepted. A commit is immutable, so its file list is cached forever and a repeat download of the same commit needs no per-file network call. Failure modes. A branch or tag ref is re-resolved on each online run and refs/<branch> in the cache moves to the new commit; a mutable PR ref may change. Limitations. Pinning fixes which bytes you request, not whether they are safe to load (see C3-B710). Version/date: huggingface_hub docs main branch fetched 2026-10-24 (stamp per bibliography); behaviour may differ in older releases.
• articleDo the 'reverse proxy' mentions in the Manager configuration page mean I can put Manager behind an inbound reverse proxy? No: they are outbound mirror variables GITHUB_ENDPOINT and HF_ENDPOINT, not listener settings.
  Problem. The configuration page uses the words reverse proxy, which could be misread as guidance for exposing ComfyUI. Conditions. Reading docs.comfy.org Manager configuration page, Environment variables table. Method. Both mentions are environment variables for outbound traffic: GITHUB_ENDPOINT (reverse proxy for GitHub access) and HF_ENDPOINT (reverse proxy for Hugging Face access); the only example redirects GitHub requests through a mirror URL. They change where Manager fetches node repositories and models from, not how clients reach ComfyUI. Failure modes. Treating them as inbound proxy guidance leaves the listener question unanswered; the page has no inbound reverse-proxy, TLS, authentication or port guidance. A third-party mirror also becomes a trust point for node code (inference: the example uses a public ghproxy mirror host). Limitations. The page does not say whether these variables are honoured by every install path or by cm-cli. Version/date. Fetched 2026-10-24T01:00:04Z.
• articleHow exactly do --listen, --port and the TLS flags behave, and is mutual TLS or any login offered?
  Problem. Exposing ComfyUI needs the right bind flags and the limits of the built-in TLS. Conditions. Direct exposure without a proxy. Method. --listen defaults to 127.0.0.1; bare --listen means 0.0.0.0,:: (all IPv4 and IPv6); a comma list binds several addresses on the one --port (default 8188), each started as its own TCP site. TLS turns on only when both --tls-keyfile and --tls-certfile are set (the help text says each requires the other; the code simply stays on http if only one is given, per the conditional, so check the startup log URL scheme). Failure modes. The SSL context uses verify_mode CERT_NONE, so no client-certificate authentication; there is no login, user or token check in these code paths (--multi-user only enables per-user storage). Port in use exits with an error. Limitations. Put authentication and rate limiting at a proxy or tunnel. Version. Commit fb3befd.
• articleWhat does --enable-cors-header actually do, and does enabling it weaken the built-in origin check?
  Problem. People add --enable-cors-header to let a web front end call ComfyUI. Conditions. Any start with the flag. Method. The flag takes one optional origin string (bare flag means '*'). Server.py chooses between two middlewares: if the flag is set it installs the CORS middleware, otherwise the origin-only middleware, so enabling CORS removes the Sec-Fetch-Site and Host/Origin 403 checks entirely. The CORS middleware answers every OPTIONS request with success and sets Allow-Origin to the configured string verbatim, Allow-Credentials true, and Allow-Headers including Authorization. Failure modes. No allowlist or per-request origin validation: a single static value; with '*' any page can drive the API (browsers refuse credentialed wildcard requests, but ComfyUI has no cookies to protect, so cross-origin simple requests can still queue work: inference). Limitations. Prefer a specific origin; put authentication in a proxy. Version. Commit fb3befd.
• articleWhich built-in guard stops a website from posting workflows to ComfyUI, and does it still work when ComfyUI is listening on a LAN address or behind a reverse proxy?
  Problem. Operators assume the Host/Origin check protects any exposed ComfyUI. Conditions. Default start (no --enable-cors-header). Method. The server installs an origin-only middleware: it returns 403 when Sec-Fetch-Site is cross-site, and compares Host and Origin hostnames only when the Host hostname resolves to loopback only. Failure modes. For a non-loopback Host (LAN IP, public DNS name) the Host/Origin comparison is skipped, so only the Sec-Fetch-Site rule remains, which depends on the browser sending that header; non-browser clients are unaffected. A proxy that rewrites Host to 127.0.0.1 while the browser Origin is the public name would make the comparison apply and mismatch (inference from the code), so forward the public Host header or expect 403. Limitations. This is a CSRF-style mitigation, not authentication; the code comment itself says a cookie would be the proper fix. Version. Source at commit fb3befd.
• articleWhat music editing tasks does MelodyFlow cover (inpainting, outpainting, instrument swap), and what is documented about its speed and availability?
  Problem. Text-driven editing of existing music, including filling a gap. Conditions. Meta research model on a 48 kHz stereo VAE latent with flow matching. Method. The paper lists editing tasks as inpainting/outpainting, looping, instrument or genre swapping, vocals removal, lyrics editing, tempo control and recording-condition changes, claims no prior approach showed high-fidelity generic style transfer across them, and evaluates with FAD, KL, CLAP similarity, edit-distance metrics and listening tests. Measured result. 64 inference steps; table latency 1.8 s (small) and 2.3 s (medium) vs 18.1 s for AudioLDM 2 (hardware not stated in the sections read). Related inpainting work listed: masked acoustic modelling for inpainting, outpainting, continuation and vamping. Limitations. At submission weights were promised, not released; verify. Version/date. arXiv 2407.03648 (2024).
• articleWhat precision (fp16, bf16, fp8) paths does ComfyUI's source assume on AMD GPUs, and which Radeon generations lose bf16?
  Problem. Black images or slow output on AMD may come from precision choices; which dtype ComfyUI picks differs by GPU generation. Conditions. ComfyUI commit fb3befd on ROCm. Method. should_use_fp16 returns True for any HIP build (torch.version.hip). should_use_bf16: for archs in the RDNA2-and-older list (gfx1030..gfx1036, gfx1010..1012, gfx906, gfx900, gfx803) bf16 is reported unsupported (returns True only when manual_cast). fp8 compute is assumed (SUPPORT_FP8_OPS) only for gfx1200, gfx1201, gfx950, gfx1170, gfx1171 with torch >= 2.7 and ROCm >= 6.4; gfx942 is excluded with a comment about an error on a nightly. The --fast option fp16_accumulation is allowed on NVIDIA or AMD. Flag help warns --fp16-vae 'might cause black images'. Example. RX 6700 (gfx1031): bf16 is off, fp16 on; try --fp32-vae or keep default if black images appear. Failure modes. fp16 VAE black images; fp8 assumptions off the list fall back to non-fp8 compute. Limitations. Lists reflect a commit; no AMD hardware r...
• articleWhat do Demucs's own README and licence say: is it maintained, what licence, and what SDR does htdemucs report?
  Problem: deciding whether to depend on Demucs v4 for stem separation. Conditions: repository facebookresearch/demucs. Method: the README states the repo is no longer maintained by its author (a fork is named for bug fixes), v4 is Hybrid Transformer Demucs (default htdemucs, trained on MUSDB HQ plus 800 extra songs), 9.00 dB SDR on the MUSDB HQ test set, 9.20 dB for the sparse-attention fine-tuned model not released; htdemucs_ft is about four times slower; htdemucs_6s adds guitar and piano. Code licence MIT (Meta). Failure modes: unmaintained code may break with new torch/torchaudio versions (inference); the 6-source model has weak piano. Limitations: the licence covers code; the README does not describe training-data terms for the 800 songs or output rights. Version/date: fetched 2026-10-23.
• articleIs there a machine-readable catalogue of official templates that lets me preflight minimum ComfyUI version, required custom nodes and download size before loading one?
  Problem. Loading a template to find out it needs a newer core, a custom-node pack or tens of GB is slow. Conditions. Reading templates/index.json from the repo (also shipped in the package). Method. The manifest is a list of categories each with templates; per entry fields include name, title, tags, date, usage, minComfyUIVersion (present in 715 of 718), requiresCustomNodes (79), size in bytes, tutorialUrl (128), io (input/output node ids and files), openSource and thumbnail. Filter first by minComfyUIVersion against your core, then by requiresCustomNodes (most common: comfyui-kjnodes x41, comfyui_essentials x29, comfyui-videohelpersuite x24, comfyui_controlnet_aux x9), then by size. Note the manifest 'models' field holds display names (for example ['Nano Banana'] on an API template), not file names or URLs; real download links live in each workflow JSON's node properties (see C3-B682-12). Failure modes. Treating manifest 'models' as downloadable files; assuming size is present for ...
• articleHow tight can machine-translated dubbing script length be to the source, and what does it cost in quality?
  Problem. Dubbing needs translation matching utterance duration. Conditions. Research on English to French, Italian, German and Spanish with a public benchmark. Method. Isometric MT trains a transformer to produce output within about plus or minus 10% of the source in character count without N-best re-ranking; a companion paper compares verbosity control methods and reports subjective tests on dubbed clips. Practical use: set a length ratio target (about 0.9-1.1 of source characters) when prompting or selecting among translations, then fit residual difference with TTS rate (C3-B792-02) and pauses. Failure modes. Length control can cost translation quality; character count is a proxy for duration. Limitations. I read only abstract pages, not numbers in the papers; results are for four language pairs. Version/date. arXiv 2112.08682 and 2110.03847. Links C3-B591-01 for the synchrony concept.
• articleHow does FaceFormer map speech to 3D face meshes, what data and licence does it need, and what are its stated limits?
  Problem. Understanding the baseline method behind many speech-to-3D-face tools. Conditions. Research use with a neutral template mesh. Method. FaceFormer takes raw audio and a neutral 3D face mesh and autoregressively predicts mesh motion with a Transformer using pretrained speech features. Code is MIT. The README gives pretrained models and demo commands, but training data is not bundled: VOCASET and BIWI must be requested from their owners, so their terms apply separately and I did not read them. Failure modes. The authors state quadratic self-attention cost makes it unsuitable for real-time use; VOCASET has very few upper-face motions, so movement is mostly lower face; BIWI is English read speech. Limitations. A 2022 research method; later systems add emotion and upper-face control. Version/date. CVPR 2022; repo last pushed 2023-08-22.
• articleCan I train Audio2Face-3D on my own character, language or style, and what do I need?
  Problem. The pretrained models are tied to a few identities and languages. Conditions. Linux or WSL2 with an NVIDIA GPU and Docker. Method. The Training Framework README says it supports prebuilt and custom models for specific characters, languages or styles; it needs paired audio and facial animation data, Docker, NVIDIA Docker and at least 6 GB VRAM, then preprocessing, training, deployment, a TensorRT engine build, and inference through Maya-ACE, SDK or NIM. A Claire example dataset (gated, demo only) shows the layout. The paper reports training times for its own networks, for example 28 and 40 minutes for the actor-specific v2.3 nets. Failure modes. Little or no capture data; the paper notes tongue data must be synthesised. Limitations. Those timings come from the authors' hardware and data; mine would differ. Version/date. Repo pushed 2025-10-15 per GitHub metadata.
• articleWhat does it take to run Audio2Face-3D inference (SDK build, GPU, streaming window), and what tuning knobs exist after inference?
  Problem. Planning local, real-time audio-to-face. Conditions. NVIDIA GPU with CUDA and TensorRT; Windows or Linux. Method. The SDK README lists building from source with CUDA 12.8 or later (below 13), TensorRT 10.13 or later, and recommends 8 GB RAM and 4 GB GPU memory. It claims faster-than-60-FPS frame generation. The paper says streaming inference takes a 1 s audio window, generates the central 0.5 s, strides 0.5 s, and uses 2 diffusion steps. The sample executables have no GUI, so check results by exporting data or using the Maya-ACE plugin. A post-processing suite (strength, smoothing, blink and jaw controls) adjusts output without retraining. Failure modes. Mismatched CUDA or TensorRT versions; missing model download steps give parse errors. Limitations. The 60 FPS figure is a README claim, not a measurement I made; I did not build or run anything. Version/date. SDK pushed 2025-08-28 per GitHub metadata.
• articleMarigold V2 in ComfyUI: what licence applies to the LoRAs versus the Qwen base, and what GPU memory does upstream state?
  Problem. The ComfyUI Marigold V2 tutorial gives no licence or memory figure. Conditions. Depth, normals or albedo estimation on a single GPU. Method. Upstream states the Marigold weights are Apache 2.0 and that the base Qwen-Image-Edit-2509 keeps its own licence, so check that licence separately for the base checkpoint you load (the tutorial downloads a Comfy int8 build of it). Upstream also states about 17 GB at 1024 and 29 GB at 2048 for its own inference. Failure modes. Reading the Comfy-Org Apache-2.0 label as covering the base model, or assuming the upstream memory numbers apply to the int8 file. Limitations. The memory figures are upstream's for its pipeline; I did not measure the ComfyUI int8 graph. I did not fetch the Qwen licence. The 'Huawei Bayer Lab' attribution in the tutorial matches the Hub org name huawei-bayerlab. Version/date. Fetched 2026-10-24.
• articleDoes TripoSplat use DINOv2 or DINOv3 in ComfyUI, and which licence applies?
  Problem. The TripoSplat tutorial labels the vision encoder DINOv2 in prose and in the download list, but the file it links is dino_v3_vit_h.safetensors. Conditions. Any agent writing install steps or licence notes for TripoSplat. Method. Trust the file name that matches the upstream card layout (clip_vision/dino_v3_vit_h.safetensors) and treat the DINOv2 wording as a documentation inconsistency. The tutorial says only 'released under an open-source license'; the upstream model card front matter states MIT. Failure modes. Writing a DINOv2 attribution or licence note (DINOv2 and DINOv3 carry different terms) from the tutorial text. Limitations. I did not check which DINO version the graph actually loads, and I did not fetch a DINOv3 licence, so I make no claim about it. Version/date. Fetched 2026-10-24.
• articleWhy does the ComfyUI SDPose workflow use RT-DETRv4 when the upstream card tells you to download YOLO11-x, and what does the MIT card not say about the Stable Diffusion v2 backbone?
  Problem. Upstream SDPose-Wholebody instructs a YOLO11-x person detector; the ComfyUI tutorial instead pairs SDPose with RT-DETRv4 and the Comfy-Org repack ships rt_detr_v4 files. Conditions. Multi-person pose or pose-map generation in ComfyUI. Method. Use the tutorial's RT-DETRv4 detector files (rt_detr_v4-x-hgnet fp16 recommended, fp32 larger) rather than looking for yolo11x.pt; the upstream repo itself carries a yolo11x.pt. Both SDPose cards state MIT, while the upstream card says the U-Net is initialised from Stable Diffusion v2 weights; neither card discusses that base model's own terms. Failure modes. Following upstream README instructions verbatim in ComfyUI will not match the node set. Limitations. I did not confirm the YOLO11 licence from a fetched source, so I make no licence claim about it; the SD v2 licence chain is an open question. Version/date. Fetched 2026-10-24.
• articleWhich Depth Anything 3 checkpoints that ComfyUI ships are commercially usable? Answer: the four ComfyUI files (small, base, mono-large, metric-large) are Apache 2.0 upstream; the CC BY-NC 4.0 variants (large, giant, nested) are not the ones the tutorial lists.
  Problem. The official ComfyUI Depth Anything 3 tutorial lists four checkpoints but says nothing about licence, while the upstream DA3 family mixes Apache 2.0 and CC BY-NC 4.0 by model size. Conditions. Use when choosing a DA3 file for commercial or client work. Method. Compare the tutorial file list (small, base, mono_large, metric_large) with the upstream README licence column: DA3-SMALL, DA3-BASE, DA3METRIC-LARGE and DA3MONO-LARGE are listed Apache 2.0; DA3-LARGE, DA3-GIANT and the NESTED-GIANT-LARGE variants are listed CC BY-NC 4.0. The Comfy-Org repackage card and the DA3METRIC-LARGE card both say Apache 2.0. Failure modes. Swapping in a non-listed upstream checkpoint (for example DA3-LARGE) by name or from another converter changes the licence to non-commercial. Limitations. Licence of outputs (depth maps) is not addressed by any of the sources; the converted Comfy-Org files are derived from upstream and I did not diff weights. Version/date. Pages fetched 2026-10-24.
• articleWhich PyMeshLab/MeshLab filters repair non-manifold geometry and duplicates, and which only select problem faces?
  Problem. Scripted repair of a batch of generated meshes. Conditions. PyMeshLab (Python bindings of MeshLab filters). Method. Detect: compute_selection_by_self_intersections_per_face (selects self-intersecting faces), compute_selection_by_non_manifold_edges_per_face. Repair: meshing_repair_non_manifold_edges (removes faces or splits vertices, so each non-manifold edge chain can become a border), meshing_repair_non_manifold_vertices (splits vertices until 2-manifold), meshing_remove_duplicate_vertices (merges same-coordinate vertices), meshing_close_holes, meshing_re_orient_faces_coherently. Failure modes. Removing smallest-area faces opens holes: follow with close-holes then re-check; self-intersections are only selected, not removed (inference: need remeshing or manual cleanup). Limitations. Parameters not exercised; no AI-mesh test run. Version and date. readthedocs latest at fetch.

More