dkackman · Nov 1, 2026 · Oct 28, 2026 · Oct 30, 2026 · Oct 31, 2026 · Oct 31, 2026
diff --git a/README.md b/README.md
@@ -10,7 +10,7 @@ two front ends on it: an **MCP server**, so Claude Code (or any MCP client)
 can author, run and inspect generations; and a **web UI** for doing the same
 by hand. A CLI sits underneath for when you want neither.
 
-**Python 3.10-3.14 | CUDA (NVIDIA) | MPS (Apple Silicon) | CPU**
+**Python 3.12-3.14 | CUDA (NVIDIA) | MPS (Apple Silicon) | CPU**
 
 <picture>
   <source media="(prefers-color-scheme: dark)" srcset="docs/img/ui-workflows-dark.png">
@@ -245,3 +245,4 @@ frame interpolation and more.
 - [Testing](docs/TESTING.md) — Running the test suite
 - [Releasing](docs/RELEASING.md) — Cutting a release from a version tag
 - [Agent Loop](docs/AGENT_LOOP.md) — What the automated implementer/tester ticket labels mean, for anyone joining an Issue
+- [Attributions](docs/ATTRIBUTIONS.md) — Code included from other projects, and the projects whose ideas dw reimplemented
diff --git a/docs/ARCHITECTURE.md b/docs/ARCHITECTURE.md
diff --git a/docs/ATTRIBUTIONS.md b/docs/ATTRIBUTIONS.md
@@ -0,0 +1,111 @@
+# Attributions
+
+diffusers-workflow is Apache-2.0. This page credits the work it builds on that
+a dependency list does not show: code that was copied in, and other projects
+whose ideas were reimplemented here. Python dependencies are in
+[requirements.txt](../requirements.txt) and [pyproject.toml](../pyproject.toml).
+Model and LoRA weights keep their own licences, and each catalog entry
+([workflows/models/](../workflows/models/), [loras/](../loras/)) names the
+repository it downloads from.
+
+## Code included in this repository
+
+Each file keeps its upstream licence notice in its header.
+
+| File | Source | Licence |
+|---|---|---|
+| [dw/tasks/rife_model.py](../dw/tasks/rife_model.py) | The RIFE interpolation network from [hzwer/Practical-RIFE](https://github.com/hzwer/Practical-RIFE) | MIT, Copyright (c) Megvii Inc. |
+| [dw/tasks/h3_latent_upscaler_model.py](../dw/tasks/h3_latent_upscaler_model.py) | The MiniMax H3 3D-convolution latent upscaler from [LBH-123-AI/Comfyui_Minimax_h3_latent_Upscaler](https://github.com/LBH-123-AI/Comfyui_Minimax_h3_latent_Upscaler) | MIT, Copyright (c) LBH-123-AI |
+| [dw/vendored/audio8_tts_model.py](../dw/vendored/audio8_tts_model.py) | The Audio8-TTS (ArkTTS) dual-AR speech model, codec and processor from [Edge0/Audio8-TTS-Realistic-0.6b](https://huggingface.co/Edge0/Audio8-TTS-Realistic-0.6b/tree/1e857568bfd4f02931aca3988ee773a44d2d26d3), ported (upstream NOTICE kept in the file header) | Apache-2.0 |
+| [dw/vendored/mmaudio_model.py](../dw/vendored/mmaudio_model.py), [dw/vendored/mmaudio_clip.py](../dw/vendored/mmaudio_clip.py) | The MMAudio flow-matching network, its sequence lengths and classifier-free guidance, and the CLIP conditioning from [hkchengrex/MMAudio](https://github.com/hkchengrex/MMAudio/tree/974010a026c731054592d8f777218bd9d85a6c24), ported (the sampler is dw's own Euler loop, and the DFN5B towers are transformers' CLIP models filled by a key map) | MIT, Copyright (c) 2024 Sony Research Inc. |
+| [dw/vendored/mmaudio_sync_model.py](../dw/vendored/mmaudio_sync_model.py) | MMAudio's Synchformer visual encoder (a MotionFormer feature extractor) and its 16-frame segmenting, from [hkchengrex/MMAudio](https://github.com/hkchengrex/MMAudio/tree/974010a026c731054592d8f777218bd9d85a6c24)'s `mmaudio/ext/synchformer/`, ported without timm, einops or omegaconf | MIT, Copyright (c) 2024 Vladimir Iashin |
+| [dw/tasks/mmaudio_decoder.py](../dw/tasks/mmaudio_decoder.py) | Loads diffusers' `MMAudioVAE` and `MMAudioVocoder`; the vocoder is [BigVGAN](https://github.com/NVIDIA/BigVGAN) (no BigVGAN code is copied here: the classes are diffusers') | BigVGAN: MIT, Copyright (c) 2024 NVIDIA CORPORATION |
+| [dw/community_pipelines/pipeline_flux_rf_inversion.py](../dw/community_pipelines/pipeline_flux_rf_inversion.py) | The diffusers community RF-Inversion pipeline, after [RF Inversion](https://rf-inversion.github.io/) by Litu Rout, Yujia Chen, Nataniel Ruiz, Constantine Caramanis, Sanjay Shakkottai and Wen-Sheng Chu | Apache-2.0, Copyright 2024 Black Forest Labs and The HuggingFace Team |
+| [dw/prompt_weighting.py](../dw/prompt_weighting.py) | Based on [sd_embed](https://github.com/xhinker/sd_embed) by Andrew Zhu | Apache-2.0 |
+
+## Model weights with their own terms
+
+Most weights are named only by their catalog entry. These are listed here as
+well, because their terms are narrower than this repository's licence and
+apply to anything generated with them.
+
+| Weights | Used by | Licence |
+|---|---|---|
+| [hkchengrex/MMAudio](https://huggingface.co/hkchengrex/MMAudio) (`weights/mmaudio_large_44k_v2.pth` and its `ext_weights/`) | `generate_sound_effects`: [dw/vendored/mmaudio_model.py](../dw/vendored/mmaudio_model.py) | [CC-BY-NC 4.0](https://creativecommons.org/licenses/by-nc/4.0/): non-commercial use only |
+| [hkchengrex/MMAudio](https://huggingface.co/hkchengrex/MMAudio) `ext_weights/synchformer_state_dict.pth` | `generate_sound_effects`, MMAudio's sync features, [dw/vendored/mmaudio_sync_model.py](../dw/vendored/mmaudio_sync_model.py) | [CC-BY-NC 4.0](https://creativecommons.org/licenses/by-nc/4.0/): non-commercial use only |
+| [kandinskylab/Kandinsky-6.0-Lite-distill-5s-Diffusers](https://huggingface.co/kandinskylab/Kandinsky-6.0-Lite-distill-5s-Diffusers/tree/4c487e8b4aeeb9059e5a77905aa2ddb59ec28f81) `audio_vae/` and `vocoder/` | `generate_sound_effects`, MMAudio's decoder, [dw/tasks/mmaudio_decoder.py](../dw/tasks/mmaudio_decoder.py) | The MMAudio VAE tensors: [CC-BY-NC 4.0](https://creativecommons.org/licenses/by-nc/4.0/), non-commercial use only, as MMAudio's weights (K6's model card states MIT; dw follows the stricter MMAudio terms). The BigVGAN vocoder: MIT. |
+| [apple/DFN5B-CLIP-ViT-H-14-384](https://huggingface.co/apple/DFN5B-CLIP-ViT-H-14-384) | `generate_sound_effects`, MMAudio's CLIP conditioning, [dw/vendored/mmaudio_clip.py](../dw/vendored/mmaudio_clip.py) | [Apple Machine Learning Research Model License](https://huggingface.co/apple/DFN5B-CLIP-ViT-H-14-384/blob/main/LICENSE): research purposes only. Apple Machine Learning Research Model is licensed under the Apple Machine Learning Research Model License Agreement. dw maps the weights into transformers' layout in memory at load time and never writes or ships converted tensors. |
+
+## Ideas reimplemented here
+
+These projects showed an approach that dw rebuilt in its own code for
+diffusers. No code, prompt text or data was copied from them; where a licence
+does not allow reuse, only the concept was taken. The issue numbers link to the
+design discussion, which names the source files consulted.
+
+### [vrgamegirl19/comfyui-vrgamedevgirl](https://github.com/vrgamegirl19/comfyui-vrgamedevgirl)
+
+A ComfyUI node pack and set of workflows for MiniMax H3 and LTX-2 film making,
+assessed in October 2026. It is source-available under a licence that does not
+allow reuse here, so only concepts were taken. It inspired:
+
+- holding the reference audio during H3 denoising, and a refine pass after the
+  latent upscale ([#598](https://github.com/dkackman/diffusers-workflow/issues/598))
+- tracked, distance-gated face repair with LTX on face crops ([#599](https://github.com/dkackman/diffusers-workflow/issues/599))
+- the music-video timing stack: beat analysis, a lyric- and beat-aware cut
+  planner, pad-then-trim shots ([#600](https://github.com/dkackman/diffusers-workflow/issues/600))
+- overlapping windows for long video sources, joined with a cosine blend
+  (`window_video`, `join_windows`; [#601](https://github.com/dkackman/diffusers-workflow/issues/601))
+- the exact size and frame-count round trip for video-to-video templates
+  (`fit_to_model`, `restore_to_source`; [#602](https://github.com/dkackman/diffusers-workflow/issues/602))
+- the finishing tasks: LUT apply and make, film grain, sharpen, tonal controls
+  in `grade` ([#603](https://github.com/dkackman/diffusers-workflow/issues/603))
+- stem separation as its own task (`separate_stems`; [#604](https://github.com/dkackman/diffusers-workflow/issues/604))
+- strength-selectable same-size LTX refine and per-segment strength decay in
+  chains ([#606](https://github.com/dkackman/diffusers-workflow/issues/606))
+- laying out a reference sheet from individual images (`ingredients_grid`; [#607](https://github.com/dkackman/diffusers-workflow/issues/607))
+- the H3 dialogue-budget warnings ([#608](https://github.com/dkackman/diffusers-workflow/issues/608))
+- checking a take's transcript against the expected lines (`check_script`; [#609](https://github.com/dkackman/diffusers-workflow/issues/609))
+- the script-to-video shot-list contract and voice bible ([#610](https://github.com/dkackman/diffusers-workflow/issues/610))
+- H3 chunk continuation in latent space ([#611](https://github.com/dkackman/diffusers-workflow/issues/611))
+- trying the Alissonerdx H3 LoRAs, which its workflows use ([#612](https://github.com/dkackman/diffusers-workflow/issues/612))
+
+### [Kablex/ComfyUI-Ref2VA-VSA](https://github.com/Kablex/ComfyUI-Ref2VA-VSA)
+
+Apache-2.0. Showed that MiniMax H3's Ref2VA attention could run block-sparse,
+building on the [FastVideo](https://github.com/hao-ai-lab/FastVideo) team's
+VSA-H3 sparse-attention work. dw's `sparse` attention backend
+([dw/sparse_attention.py](../dw/sparse_attention.py),
+[#931](https://github.com/dkackman/diffusers-workflow/issues/931)) runs on
+[comfy-kitchen](https://pypi.org/project/comfy-kitchen/)'s Sol-Attn kernel, a
+dependency.
+
+### [mo-akkakk/ComfyUI-MiniMaxH3-TemporalWindow-SpatialTiles](https://github.com/mo-akkakk/ComfyUI-MiniMaxH3-TemporalWindow-SpatialTiles)
+
+MIT. The tiled low-sigma refine after an H3 upscale (`refine_tiles`,
+[#1043](https://github.com/dkackman/diffusers-workflow/issues/1043)), and the
+approach of the test that checks a one-segment H3 chain is bit-identical to a
+native run on a tiny random-weight model on CPU
+([#1046](https://github.com/dkackman/diffusers-workflow/issues/1046)).
+
+### [harlixay7/Kinoloom](https://github.com/harlixay7/Kinoloom)
+
+MIT. Its seamless-chain research gave `analyze_seams` its saturation,
+lightness, contrast and detail measures
+([#1031](https://github.com/dkackman/diffusers-workflow/issues/1031)), and the
+`script-to-video` skill its guidance on re-rolling a chained segment whose seam
+fails ([#1034](https://github.com/dkackman/diffusers-workflow/issues/1034)).
+
+### [r34vtraining/Longshot_Studio](https://github.com/r34vtraining/Longshot_Studio)
+
+A ComfyUI studio for making long MiniMax H3 films shot by shot, assessed in
+October 2026. It has no licence, so only concepts are taken. Its ideas are
+filed as dialogue/effects stem separation
+([#1091](https://github.com/dkackman/diffusers-workflow/issues/1091)), an H3
+face refine with the audio held
+([#1092](https://github.com/dkackman/diffusers-workflow/issues/1092)), cutaway
+and lead-in joins ([#1093](https://github.com/dkackman/diffusers-workflow/issues/1093)),
+rendering the odd shots of a long take first and bridging the gaps
+([#1094](https://github.com/dkackman/diffusers-workflow/issues/1094)), and
+finding phrase starts from the centre of the stereo mix
+([#1095](https://github.com/dkackman/diffusers-workflow/issues/1095)).
diff --git a/docs/DEPENDENCIES.md b/docs/DEPENDENCIES.md
@@ -16,7 +16,7 @@ source ./activate
 .\venv\scripts\activate
 ```
 
-The install scripts handle everything: Python version detection (3.10-3.14), virtual environment creation, PyTorch, diffusers, and all dependencies.
+The install scripts handle everything: Python version detection (3.12-3.14), virtual environment creation, PyTorch, diffusers, and all dependencies.
 
 ### Manual Installation
 

diff --git a/docs/MCP.md b/docs/MCP.md
@@ -424,11 +424,18 @@ one phase, so two polls otherwise come back identical: read `denoise_step`
 moving (slow but healthy) against a `denoise_step` that is a number and
 stays put while `seconds_since_event` climbs (nothing is happening). A null
 `denoise_step` under `generating` is neither - it is the lead-in the
-pipeline runs before the loop, encoding the prompt and every reference, with
-nothing emitted, so silence there is expected. Its length follows what it
-encodes: ~90 s on MiniMax H3 for a prompt with an image or audio reference,
-~10 min once a *video* reference is among them (measured: 629 s for one 5 s
-960x544 clip on an RTX 3090). `get_job_events` names the block it is in
+pipeline runs before the loop, encoding the prompt and every reference and
+then the first denoise pass, with nothing emitted, so silence there is
+expected. On MiniMax H3 the first pass dominates the lead-in, not the encode,
+and it varies by card. The reference encode (`vae_encoder`) is about 6 s for
+an image reference and about 50-53 s for a 5 s 960x544 video reference;
+reference tokens make the first pass slower. Measured on lem: an image
+reference's lead-in was 187 s (job `f90a041af26a`: `timing.encode_s` 26 s,
+first pass 161 s; of the encode, `vae_encoder` was 6 s); the same image +
+video job took 508 s on cuda:1 (`f5aefcdd5e26`: `encode_s` 86 s, first pass
+422 s; `vae_encoder` 59 s) and 165 s on cuda:0 (`6961876f11d3`: `encode_s`
+79 s, first pass 86 s; `vae_encoder` 56 s). Plan for several minutes, up to
+~8.5 min measured (508 s, `f5aefcdd5e26`), with a video reference. `get_job_events` names the block it is in
 while that runs - a `log` line per top-level block of a modular pipeline
 (`MiniMaxAI/MiniMax-H3: vae_encoder`), which is the difference between
 silence and knowing it is encoding the reference. And once the counter is a number the gaps