You need to agree to share your Hugging Face information to access this model
This repository is publicly accessible, but you have to accept the conditions to access its files and content.
Access is subject to the upstream MiniMax H3 Community License Agreement and Acceptable Use Policy in LICENSE. Its territory restrictions originate with MiniMax, not with an additional license imposed by Bunsei. The EU, UK, Republic of Korea and USA are excluded territories under that upstream license. Do not request access if you or your intended use fall outside the Applicable Territory. This experimental conversion has no ongoing maintenance commitment.
Log in or Sign Up to review the conditions and access this model content.
MiniMax H3 Ref2VA FP8 for Radiance
中文模型卡 · Experimental engine and documentation
An experimental, preconverted .rad container for the MiniMax H3 Ref2VA path in Bunsei's Radiance full-reference fork. It packages an existing pruned FP8-scaled checkpoint for that runtime. It is not a newly trained model and is not an official MiniMax release.
Use the matching fork for image, video, audio and mixed-reference conditioning. A .rad file alone does not add full-reference support to an arbitrary upstream Radiance build. It cannot be loaded directly by ComfyUI, Transformers or Diffusers.
Artifact and provenance
| Item | Value |
|---|---|
| File | minimax-h3-fullref-experimental-fp8.rad |
| Size | 26,863,880,192 bytes (26.86 GB; 25.02 GiB) |
| SHA-256 | 6d886c33067f81a3e6cc435351b42702b64a2c0ddfd81aa1387b87f4bd46f1a8 |
| Original model | MiniMaxAI/MiniMax-H3 |
| Ref2VA checkpoint | Comfy-Org/MiniMax-H3, minimax_h3_ref2va_pruned_fp8_scaled.safetensors |
| Container base | StillDeadcode/minimax-h3-fp8 |
| Runtime source snapshot | 41ded9231759d28ef095338aa103ce997be73a85, branch h3-fullref |
The Ref2VA source checkpoint SHA-256 is f86f2f79ebd2d76eb8eeb46091e83982e6ff51d255747e7b16e92834b392b8e9; recorded source revision: e5eb578a89295337b8ff433a035929ce0279e0b6.
The importer replaces the refiner/DiT components, adapts scale layout while preserving the source Ref2VA FP8 codes and scales, and carries over the base container's decoders and vocabulary. Existing FP8 quantization and pruning were inherited from the upstream conversions. This repack adds no training, fine-tuning, calibration or further quantization. The container includes mixed-precision components; “FP8” does not mean every tensor or operation is FP8.
No acceleration LoRA is merged. No user images, reference videos, music, prepared latents or credentials are embedded. The tested imported weights were not subsequently adjusted. See MODEL-MANIFEST.json, NOTICE, and the original base-model notice.
Runtime requirements
The demonstrated configuration is Linux x86-64, two AMD Radeon AI PRO R9700 GPUs (32 GiB each, gfx1201), tensor parallelism with an operational PCIe P2P path, and host memory offload. Two cards do not behave as a single unified 64 GiB GPU: some allocations are replicated and CPU memory is still used.
The accompanying experimental runtime was exercised on Ubuntu 26.04.1 / glibc 2.43. That is the prebuilt runtime's tested platform, not an intrinsic operating-system requirement of the weights. Other hardware, driver combinations and single-card deployment are not validated here. The full-reference benchmark reached roughly 34.6 GB of process resident memory in addition to GPU allocations; allow host-memory headroom. Runtime ABI is 28: build the engine and its plugins together.
Raw-media preprocessing still requires external ComfyUI encoders. The native Radiance refiner/DiT/decoders perform generation after preparation. This file is not the complete raw-media encoding stack.
Download and verify
After accepting this repository's access terms, authenticate with a Hugging Face account that has access:
hf auth login
hf download Bunsei/MiniMax-H3-Ref2VA-FP8-Radiance --local-dir /path/to/h3-rad
sha256sum /path/to/h3-rad/minimax-h3-fullref-experimental-fp8.rad
Keep all accompanying licenses and notices when redistributing the model. This download is already converted; the source Ref2VA safetensors file is only needed to repeat the import or run a separate ComfyUI comparison.
Configure and run
Follow the fork's English README or Chinese README for dependencies, ROCm setup and full deployment details.
Build a matching engine with a suitable HIP toolchain, CMake, Ninja, FFmpeg development libraries and the remaining dependencies documented in the fork:
git clone --branch h3-fullref https://codeberg.org/Bunsei/radiance-h3-fullref.git
cd radiance-h3-fullref
git checkout 41ded9231759d28ef095338aa103ce997be73a85
cmake -S . -B build -G Ninja -DRAD_WITH_HIP=ON -DRAD_GPU_TARGETS=gfx1201 \
-DRAD_WITH_FFMPEG=ON -DRAD_BUILD_TESTS=ON -DCMAKE_BUILD_TYPE=Release
cmake --build build --target radiance minimaxh3 libref libquant libr4d
Prepare a compatible ComfyUI environment (v0.40.0 was tested) and download its separate encoder weights into:
ComfyUI-Models/
text_encoders/qwen3vl_32b_minimax_h3_nvfp4_awq.safetensors
vae/minimax_h3_video_vae_fp16.safetensors
vae/minimax_h3_audio_vae_fp32.safetensors
These are available from Comfy-Org/MiniMax-H3. Observe their respective upstream terms. Then prepare an image-reference request; replace every example path with your own:
export COMFY=/path/to/ComfyUI
export MODELS=/path/to/ComfyUI-Models
export MODEL=/path/to/h3-rad/minimax-h3-fullref-experimental-fp8.rad
export CASE=/path/to/new-case
ROCR_VISIBLE_DEVICES=0 "$COMFY/.venv/bin/python" scripts/h3_fullref.py prepare \
--comfy "$COMFY" --models "$MODELS" --work "$CASE" \
--width 832 --height 480 --length 124 --steps 20 --seed 145 \
--reference-max-area 344064 \
--image /path/to/authorized-reference.jpg --prompt /path/to/prompt.txt
The prompt can refer to <Picture 1>. Repeat --image, --video, or --audio for multiple references; --video-audio VIDEO AUDIO pairs a video with its soundtrack. Reference tags are <Picture 1>, <Video 1>, <Audio 1>, and so on. Inspect the generated inputs.json and its packed_rows before generation. prepared.pt retains the prepared inputs for reproducibility; request.json is consumed by the native engine. Do not publicly redistribute those files if they encode restricted reference material.
Release preprocessing GPU allocations before starting native generation. From the fork's root, run this TP2 helper with a new output directory:
export RUN=/path/to/new-run
ROCR_VISIBLE_DEVICES=0,1 HSA_NO_SCRATCH_RECLAIM=0 HSA_SCRATCH_SINGLE_LIMIT=67108864 \
python3 scripts/h3_native_benchmark.py \
--engine build/bin/radiance --model "$MODEL" \
--radiance-home build/radiance_home --request "$CASE/request.json" \
--output "$RUN" --port 8894 -- \
--gpu-headroom-mib 2048 --host-pool-mib 12288 --p2p on \
--vision-tower-placement ram --kernels libr4d --placement layer_offload
The helper starts and stops its own server. --host-pool-mib is a per-rank pool setting, not a total process-memory ceiling. Inspect the output video, engine log, watchdog status and actual process exit; a controller summary alone is insufficient to establish success. The sampling watchdog limits each step to 600 seconds; it cannot guarantee recovery from a driver hang.
--steps 20 requests 20 denoiser evaluations (21 scheduler points including terminal zero). Native 124 frames at 24 fps span about 5.17 seconds; trim to 120 frames if exactly 5 seconds is needed. For the demonstrated full-reference case, use 243 native frames, an image, the motion video and standalone music; --pad-reference-video holds its final source frame when needed. Output trimming is separate from generation and must not change the reference motion speed.
Limits of this experimental snapshot
- Output width and height must be multiples of 32, at least 256 each, aspect ratio at most 4:1, and area at most 1,032,192 pixels (
1344 × 768). This snapshot does not support 2 MP output. - Native rate is 24 fps. Frame count is
17*n+5, withn=7..21: 124–362 frames, roughly 5.17–15.08 seconds. - Text, reference and target sequences share a 119,304-row packed budget, with separate ceilings of 16,384 presentation tokens and 65,536 reference rows.
- At most 12 reference items overall, with caps of 9 images, 3 videos and 3 standalone audio references. These are validation caps; simultaneous maximum counts can exceed the packed-row or memory budget.
- Prepared full-reference conditioning cannot be combined with the first-frame path, CFG, upscaling or FastH3 sparse attention in this snapshot.
- References guide generation; exact frame-by-frame pose, finger placement, identity and audio synchronization are not guaranteed.
- The matching fork supports acceleration-LoRA composition, but no LoRA is included, merged, or used in the benchmark below. Compatibility does not establish a quality or speed guarantee for arbitrary LoRAs.
Measured examples
Two completed native runs on 2 × R9700, 832 × 480 (0.39936 MP), 20 denoiser evaluations, seed 145, Euler, CFG 1. These are particular workloads, not a general performance promise:
| Workload | Native frames | Sampling | Decode | Longest sampling step |
|---|---|---|---|---|
| One identity image + 9.53 s motion video + music | 243 | 479.8 s | 13.2 s | 24.10 s |
| One image + text | 124 | 66.5 s | 6.5 s | 3.35 s |
Preprocessing (145.2 s and 25.0 s respectively), startup, muxing and final trimming are excluded from the sampling/decode columns. The full-reference delivery was trimmed at original speed to 9.533 s; the image-only delivery to 5 s. No conclusion about RTX 5090 performance or visual parity follows from these timings. Original reference media are not included in this repository.
Credits, development and maintenance
MiniMax created the model. Comfy-Org and StillDeadcode supplied the upstream conversions and container components. Radiance and all upstream contributions retain their original authorship and notices.
The downstream implementation, porting, optimization, debugging, test tooling and documentation were carried out by OpenAI Codex (vibe coding). Bunsei provided the requirements, hardware and reference assets, and published the experimental snapshot. This statement applies to the downstream additions, not upstream code or model training. Bunsei does not commit to ongoing development, issue triage, replies, fixes, compatibility updates or merging this branch upstream.
License and access conditions
The weights are governed by the MiniMax H3 Community License Agreement, including its Acceptable Use Policy. They are not relicensed under the engine's Apache-2.0 license. The inherited Qwen3-VL and latent-upscaler components have separate licenses in licenses/.
The license excludes the European Union, United Kingdom, Republic of Korea and United States of America from its Applicable Territory. Use and distribution must comply with the complete agreement, including downstream recipient obligations. Commercial use above the agreement's USD 20 million annual-revenue threshold requires separate authorization from MiniMax. This summary does not replace the license.
Downloads require agreement to the upstream license and confirmation of eligible territory. These territory terms are imposed by MiniMax's original license; Bunsei has not added a separate territorial license. The optional Hub EU IP-block metadata flag is not enabled. Removing that platform banner does not waive any upstream license term or expand the rights granted upstream. No new rights to third-party reference images, videos, music, characters or outputs are granted by this repository. Use authorized material and identify generated content appropriately.
- Downloads last month
- -
Model tree for Bunsei/MiniMax-H3-Ref2VA-FP8-Radiance
Base model
MiniMaxAI/MiniMax-H3