Loading
Du hast einen ComfyUI-Workflow? Erstelle direkt im Browser eine verständliche Workflow-Dokumentation.Workflow verständlich machen.Jetzt ausprobieren ↗
Critical fix. If you used the long-form MEMORY workflow in v1.1 or v1.2, it could not have worked. Update the node pack and restart ComfyUI fully — a browser refresh is not enough.
Laufzeitprofil
Quellenbeschreibung
Critical fix. If you used the long-form MEMORY workflow in v1.1 or v1.2, it could not have worked. Update the node pack and restart ComfyUI fully — a browser refresh is not enough.
The bug
H3MultishotMemorySampler called vae_decode_audio() without importing it, raising a NameError at audio decode — after all sampling had finished. So it destroyed a completed render rather than failing fast, which is the worst possible moment to fall over.
Checking every commit that has touched that file, the import appears exactly once in all of them, in the other sampler. This node had never worked in any released version. It shipped broken in v1.1 and again in v1.2.
Reported with a correct diagnosis and the correct patch by Nebuluss in issue #1 . Fixed and verified with a real two-shot render — 247 frames, audio stream present, and the embedded graph confirms the memory sampler itself produced it.
KI-generierter Kommentar
KI-generierte Erklärung auf Grundlage von Quellen- und Konfigurationsdetails. Vorschläge sind klar gekennzeichnet.
This MiniMax-H3 workflow turns a story prompt, optional opening image, and optional voice clip into one continuous multishot video with sound.
The download includes three ComfyUI workflows: H3_Keyframes.json, H3_Multishot_AIO.json, and H3_Multishot_MEMORY.json.
The multishot path joins several shots into one continuous take, with audio seams crossfaded.
Type a story or load a prompt file; for hand-written shots, paste raw JSON with each shot separated by --- lines.
You can add an opening image for shot one and a short voice clip to guide the speaker’s voice.
Use ComfyUI with the MiniMax-H3 Multishot node pack, a MiniMax-H3 checkpoint, and a text encoder; the VAEs are preset.
If you used the MEMORY workflow in v1.1 or v1.2, update the node pack and fully restart ComfyUI; a browser refresh is not enough.
Vorschlag · nicht geprüft
The listed VAEs are minimax_h3_audio_vae_fp32.safetensors and minimax_h3_video_vae_fp16.safetensors.
Other named files include MiniMax-H3-fl2va-Q5_1.gguf, MiniMax-H3-encoder-Q4_K_M.gguf, and MiniMax-H3-encoder-Q5_K_M.gguf; check which encoder filename is listed for the workflow you choose.
Vorschlag · nicht geprüft
The keyframe workflow supports up to six anchors placed by fraction or frame index.
For hand-written prompts, set the writer to passthrough (raw JSON, skip LLM) and separate shots with --- lines.
Use stereo for reference audio: a mono clip can crash the sampler.
Vorschlag · nicht geprüft
Muss das bearbeitet werden?
Melden Sie sich an, um eine Änderungsanfrage zu senden.
Quellen
1 QuelleQuellenauszüge
1 AuszugSource context: 10753 downloads · Type Workflows · Base model MiniMax H3
Also in this version: H3 Free Text Encoder
Unloads the text encoder before the DiT loads. This matters most for reference-to-video, where the Qwen3-VL encoder (~16.5 GB) and the ref2va DiT (~25.9 GB) together oversubscribe a 32 GB card. Measured on an RTX 5090, ref2va with one reference image at 124 frames:
encoder resident, DiT fully loaded: 90+ min, killed, no output encoder evicted + --reserve-vram 7: 12.3 min, success
Both are needed. Eviction alone still let the DiT load completely at 25.2 GB with nothing left for activations.
The diagnostic that found this was power draw, not utilisation . The GPU reported 98% "utilisation" at 146 W — a 5090 doing real diffusion pulls 400 W+. 98% at 146 W is a card spinning on memory transfers. With headroom it runs at 410 W. Full residency is not the goal; total residency under the card size is.
What changed in how this gets released
An import that survived two releases means the process failed, not just the code. Two checks now gate every package:
Provenance, not existence. My evidence file claimed the MEMORY workflow had a render behind it — pointing at a file actually produced by the other sampler. The gate confirmed the file existed and never asked what made it. It now requires the workflow's own node class to appear in the render's embedded ComfyUI graph.
A static import check over every shipped .py , so a call that parses fine but NameError s at runtime cannot ship again.
Everything from v1.2 is still here
Keyframes at any position — up to 6 anchors by fraction or frame index, not just first/last. Measured: an anchor requested at frame 121 landed at frame 122 of 243, reached by continuous motion with no cut.
H3 Condition Strength and H3 Reference Audio (stereo guard — a mono reference clip crashes the sampler with an unhelpful shape mismatch).
Image-to-video, the ~4x encoder-eviction speedup for multishot, crossfaded audio seams, and the self-repairing script parser.
Links
Node pack on GitHub
MiniMax-H3 GGUF (the DiT)
MiniMax-H3 Text Encoder GGUF
VAEs
Support
Everything I publish is free and stays free. If it saved you a night of debugging, tips keep the 5090 warm: Ko-fi · GitHub Sponsors · Liberapay .
Hardware and download-size estimates are not available yet.
Anforderungen
minimax_h3_audio_vae_fp32.safetensors
Datei ungeprüftVAE · SAFETENSORS · Unknown
minimax_h3_video_vae_fp16.safetensors
Datei ungeprüftVAE · SAFETENSORS · Unknown
MiniMax-H3-encoder-Q4_K_M.gguf
Nicht aufgelöstText encoder · Unknown
MiniMax-H3-fl2va-Q5_1.gguf
Nicht aufgelöstCheckpoint · Unknown
