Abiray

/

Minimax-H3-nvfp4-INT4-INT8-Convrot

like 215

Use either of these workflows: https://huggingface.co/Abiray/MiniMax-H3-GGUF/blob/main/minimax_fl2v_gguf_workflow.json or https://huggingface.co/Abiray/MiniMax-H3-GGUF/blob/main/minimax_ref2va_gguf_workflow.json. For the main GGUF model, you can use https://huggingface.co/Abiray/MiniMax-H3-GGUF/blob/main/unet/MiniMax-H3-FL2VA-Q4_K_M.gguf or pick the Ref2VA model from my repo. Next, load the GGUF text encoder (https://huggingface.co/Abiray/MiniMax-H3-GGUF/blob/main/text_encoders/qwen3vl_32b_minimax_h3-Q4_K_M.gguf) along with the audio VAE and video VAE. You can grab the VAEs from either my repo or the official ComfyUI repo—both will work perfectly. Just make sure you have ComfyUI-KJNodes, ComfyUI-Manager, and ComfyUI-GGUF installed, and update ComfyUI to the latest version. If you encounter any settings like multipliers or resize options, simply tweak or bypass them, and you'll be good to go with zero issues.

Generated using MiniMax_H3_FL2VA_pruned_mixed_int4_int8_convrot.safetensors

[INFO] Requested to load MiniMaxH3VideoVAE [INFO] Model MiniMaxH3VideoVAE prepared for dynamic VRAM loading. 4965MB Staged. 0 patches attached. Force pre-loaded 128 weights: 348 KB. [INFO] Found quantization metadata version 1 [INFO] Detected mixed precision quantization [INFO] Using mixed precision operations [INFO] Native ops: float8_e4m3fn, convrot_w4a4, int8_tensorwise, float8_e5m2 , emulated ops: nvfp4, mxfp8 [INFO] model weight dtype torch.bfloat16, manual cast: torch.bfloat16 [INFO] model_type FLOW [INFO] Requested to load MiniMaxH3 [INFO] Unloaded partially: 15385.37 MB freed, 0.00 MB remains loaded, 3710.90 MB buffer reserved, lowvram patches: 0 [INFO] Model MiniMaxH3 prepared for dynamic VRAM loading. 15162MB Staged. 0 patches attached. Force pre-loaded 210 weights: 1175 KB. 100%|██████████████████████████████████████████████████████████████████████████| 20/20 [10:55<00:00, 32.79s/it] [INFO] Requested to load MiniMaxH3AudioVAE [INFO] Model MiniMaxH3AudioVAE prepared for dynamic VRAM loading. 576MB Staged. 0 patches attached. Force pre-loaded 401 weights: 539 KB. [INFO] Model MiniMaxH3VideoVAE prepared for dynamic VRAM loading. 4965MB Staged. 0 patches attached. Force pre-loaded 128 weights: 348 KB. [INFO] Prompt executed in 00:12:36 0.4/24fps

I just used one image for this, but if you use two detailed ones as your first and last frames, the final video will look way sharper even at lower resolutions.

I just did this using https://huggingface.co/Abiray/MiniMax-H3-GGUF/blob/main/unet/MiniMax-H3-FL2VA-Q4_K_M.gguf.
It is much better regarding both audio and video but I feel like something still wrong in the animations. Thank you so much for your efforts and answering.. I am afraid that may be something is wrong in comfyUI itself.