Updated: Nov 4, 2026
toolMiniMax H3 Image Studio – 4K (Ref2VA + Turbo LoRA + SeedVR2)
Text-to-image / reference-based image generation with MiniMax H3, using the video model as a high-quality still generator. Includes an optional SeedVR2 4K refine pass.
Features
- 8-step Turbo sampling (Pika H3 Flex Turbo LoRA)
- Up to 9 reference images, each in its own group, easy to mute/unmute
- Reference ON/OFF switch, works as plain text-to-image when disabled
- Resolution presets with aspect ratio selection (up to 8 MP)
- Automatic best-frame selection from the decoded latent
- Optional SeedVR2 (7B GGUF) detail upscale to 4096 px with before/after comparer
- Separate Save nodes for original and upscaled images
How to use
1. Write your prompt in the "Text (Prompt)" node: scene, subject, lighting, style. No camera or sound directions needed, this is a still image.
2. With references: unmute the "Ref Img" nodes you need, load your images and set the switch to true. Without references: set the "Image Reference (On & OFF)" switch to false.
3. Pick aspect ratio and resolution in H3ImageResolutionPreset.
4. Run. If you don't need the upscale, mute the SeedVR2 group.
Required models
Diffusion model: minimax_h3_ref2va_fp8_e4m3fn.safetensors (33.3 GB)
Text encoder: qwen3vl_32b_minimax_h3_nvfp4_awq.safetensors (14.9 GB)
VAE: minimax_h3_video_vae_int8_convrot.safetensors (2.8 GB)
LoRA: pika_h3_flex_turbo_8step_hd_ref2va.safetensors (3.9 GB)
Upscale (optional): seedvr2_ema_7b-Q4_K_M.gguf + ema_vae_fp16.safetensors
Links:
https://huggingface.co/Comfy-Org/MiniMax-H3
https://huggingface.co/Pika-Labs/Pika-H3-Flex-Turbo-LoRA
https://huggingface.co/dimitarx/MiniMax-H3-fp8
Note: the VAE from the PrunaAI repo has a different file name. Rename it to match the workflow or select it in the loader.
Custom nodes
- ComfyUI-MiniMax-H3-Image-Studio (astropuzzo)
- rgthree-comfy
- comfyui-zip-ties
- ComfyUI-SeedVR2_VideoUpscaler (for the upscale group)
Tips
- Default is 8 MP, which is heavy. If you run out of VRAM, lower the preset to 4 MP or less. SeedVR2 can still bring it up to 4K.
- SeedVR2 uses block swapping to save VRAM: slower, but works on smaller cards.
- At 8 MP the input is already near 4K, so SeedVR2 acts mostly as a detail/refine pass. The Image Blur node before it is optional, bypass it for maximum sharpness.
- The PhotoReal LoRA in the Power Lora Loader is optional. Replace it with your own or disable it.
- Update ComfyUI to the latest version before use.
Credits
MiniMax H3 / Comfy-Org, Pika Labs (Turbo LoRA), astropuzzo (Image Studio nodes), rgthree, SeedVR2 authors.
Tested on: [RTX 3060 / Lastest ComfyUI version]. Generation time: [500-550 sec @ 4K + SeedVR2 Refine].

