sabi123456/MosaiChunk-H3-Conditioning

MosaiChunk H3 Conditioning for Pruned INT8 FL2VA / REF2VA Models (Experimental)

Converted conditioning weights for the MosaiChunk H3 Conditioning ComfyUI node. This 174.6 MB file lets users reuse an existing compatible pruned/quantized H3 checkpoint without repeating the one-time, approximately 26.14 GB streamed conditioning conversion.

This file is not a standalone video model, an ordinary LoRA, or the MosaiChunk router. It needs the custom node and the separate RAVEN streaming adapter. Downloading this file alone does not enable memory in an ordinary unmodified sampler.

Intended for compatible pruned, curve-form MiniMax H3 checkpoints, including INT8 FL2VA and experimental INT8 REF2VA/all-in-one models. The user checkpoint targeted during development is Minimax-h3_Singularity_ref2va_Pruned_v1.3_int8.safetensors. REF2VA support is an experimental adapter transfer; trained video quality and identity consistency have not been established.

Custom node and installation: Minimax-H3-Memory-Cutom-Node

Download and install

1. Install the custom node from the GitHub repository above.
2. Download raven-fl2va-conditioning.safetensors from this repository's Files tab.
3. Place it in ComfyUI/models/mosaichunk/.
4. Obtain the RAVEN streaming preview adapter and place minimax_h3_raven_streaming_lora_4nfe_preview.safetensors in ComfyUI/models/loras/.
5. Obtain the MosaiChunk T2V router files and place model.safetensors and config.json in ComfyUI/models/mosaichunk/t2v/.
6. Restart ComfyUI. Select the converted conditioning and mandatory RAVEN adapter on MosaiChunk H3 Conditioning. Keep your own UNET loader and other LoRAs on the MODEL chain. Do not apply the mandatory RAVEN adapter a second time through a normal LoRA loader.

Required separately: a supported existing pruned H3 generator, original FL2VA pruned conditioning baseline, H3 text encoder, video/audio VAEs and the usual H3 conditioning workflow. See the pruned-model instructions. The full BF16 checkpoint is not required for this download or to load this conditioning file.

Usage and scope

Use generation_mode=auto, conditioning_base_unet=auto when the original standard FL2VA pruned baseline is installed, 4 steps and CFG 1. The node hosts its RAVEN chunk-consistency recipe inside existing sampler workflows. It uses the supplied seed/noise and step count; it replaces the selected numerical solver and sigma grid.

Text-to-video memory and experimental first/last-frame/reference memory are implemented in the accompanying node. REF2VA transfer and image-conditioned use reuse a T2V-trained router/adapter without retraining. They are not claims of trained identity consistency, endpoint accuracy or equivalent quality on arbitrary H3 finetunes.

The conditioning file does not contain the history bank. History is created anew during each generation and released afterward. Model weights, VAEs, attention workspace and retained history all need additional system/GPU memory. This release does not establish peak memory, generation speed or a completed trained GPU benchmark. SSD storage and Spectrum/Sage compatibility are separate node-development work and are not included or promised by this weight release.

Contents and integrity

• raven-fl2va-conditioning.safetensors: 174,613,456 bytes, 103 tensors (the preserved 1025×8 time table plus weights/biases for 51 modulation projections).
• raven-fl2va-conditioning.json: conversion report and recipe.
• conditioning-coordinate-validation.json: original-coordinate rebasing measurements.
• PROVENANCE.json, SHA256SUMS, licenses and NOTICE.

Conditioning SHA256: 44b8ad251538a6d74fcbc6cc5930016e842b6cd4499b5df7c69927f34cfefb46.

Conversion evaluates the adapted time curve, folds timestep/modulation updates into rank-eight pruned coordinates and preserves the original baseline table byte for byte. The original report compares converted projections at probe timesteps against BF16 RAVEN; measured relative L2 differences are approximately 0.0017–0.0020 and include BF16 rounding. This is a numerical conversion check, not video-quality validation. Later coordinate rebasing is documented separately in the supplied reports.

Attribution and licenses

This is an unofficial modified artifact. Credit belongs to MiniMax for H3, the RAVEN authors for the streaming adapter, the MosaiChunk authors for memory routing, Comfy-Org for source packaging and YanzuoLu for the ComfyUI RAVEN backend. The linked custom-node repository supplies conversion/integration code.

The upstream MiniMax H3 terms and RAVEN CC BY-NC 4.0 terms remain applicable. See LICENSE.md, both complete license texts and NOTICE. MiniMax's grant includes territorial and acceptable-use restrictions; RAVEN contributions are noncommercial. Public hosting does not grant unrestricted worldwide or commercial use. No affiliation or endorsement by upstream authors is implied.