After many tests, this is the first and WIP version of what I currently use for Ref2VA generations.
Generate LOW RESOLUTION in a first Pass, then upscale and refine when happy with the low quality resolution results.
You can generate and see if audio etc works in low Quality (even without upscaling), and still enhance the same seed and video with the 2nd Pass of the workflow.
My workflow is optimized (WIP) for 5070ti with 16GB VRAM and 32GB SysRAM.
This is WIP so please give me some feedback how it works, what you struggle with and what you think would help others. It's very basic atm, but resulted the best speed without additional nodes etc so far, especially the First Block Cache node is helping a lot with non-Turbo 20 Step generations.
Features:
Add (currently) up to 3 image references.
Use 20 base model or 8 step Turbo Lora generation via switch
Enable or disable the high-resolution upscaling and refinement in the 2nd Pass.
Keep the same seed to produce small videos first to check for audio quality and composition before enabling the 2nd pass to upscale to higher resolution
Add your desired Loras, individually to each pass to ensure the Lora weights are applied accordingly, first pass is Ref2VA, most Loras work best with lower strengths but feel free to play around. The 2nd pass uses a FLF2VA model, use Loras here to add details for the high resolution.
MODELS
Singularity Ref2VA model: https://huggingface.co/WarmBloodAban/Minimax-h3_Singularity/blob/main/Minimax-h3_Singularity_ref2va_Pruned_v1.3_int8.safetensors
Base FLF2VA / T2V model: https://huggingface.co/Comfy-Org/MiniMax-H3/blob/main/diffusion_models/minimax_h3_fl2va_pruned_int8_convrot.safetensors
Text Encoder for 50xx Series Nvidia GPUs: https://huggingface.co/Comfy-Org/MiniMax-H3/blob/main/text_encoders/qwen3vl_32b_minimax_h3_nvfp4_awq.safetensors
Text Encoder for older GPUs: https://huggingface.co/Comfy-Org/MiniMax-H3/blob/main/text_encoders/qwen3vl_32b_minimax_h3_int8_convrot.safetensors
Audio VAE: https://huggingface.co/Comfy-Org/MiniMax-H3/blob/main/vae/minimax_h3_audio_vae_fp32.safetensors
Ref2VA Turbo Lora: https://huggingface.co/Kijai/MiniMax-H3-experimental/blob/main/loras/MiniMax-H3-Ref2VA-Acc-8Step_comfy.safetensors
Taomate 3 Step Turbo Lora: https://huggingface.co/drbaph/MiniMax-H3-Turbo-Lora-ComfyUI/blob/main/experimental/minimax_h3_taomate_fl2va_3step_ema_comfyui.safetensors
H3 Latent Upscaler: https://github.com/LBH-123-AI/Comfyui_Minimax_h3_latent_Upscaler
Custom Nodes
Description
First official version for testing and feedback from you guys
FAQ
Comments (11)
nice cant wait to test it.
so far ive been using my own version of it based on your T2V workflow .
Thanks! I really couldn't find much more to optimize for my own system, I bet there are some nodes etc. that can boost speed even more for others. Please let me know what you find out!
@zkyt trust me if something works for my 8GB VRAM and 32gb RAM i bet it will work for everyone lmao ill post results as soon as i test it .
Hi, I am seeing this error for some reason.
# ComfyUI Error Report ## Error Details - Node ID: 125 - Node Type: SamplerCustomAdvanced - Exception Type: RuntimeError - Exception Message: RuntimeError: The size of tensor a (96) must match the size of tensor b (3072) at non-singleton dimension 0
@yajukun are you using the same models as in the WF?
@zkyt Yes, most of the models were auto found on my drives, the only ones I had to navigate to were the Loras because I have a diff folder structure. It could be because I am on 1 version back on Comfy or maybe I have some conflicting custom nodes loaded. This is a well used comfy load I am using and I already have many nodes disabled because they conflicted with others. No worries. I will reload a clean comfy soon and try again. Thanks!
@yajukun best of luck! Keep me posted if you find out what's causing it.
@zkyt FYI, I reloaded everything and got it to work. I am having a conflict with the PowerLoaderV2 and being able to drag/drop images but I have a workaround for that. I am starting to wonder if I need multiple Comfy installs seperated for different models? Anyway, I am testing the WF now but it looks good so far. Thanks!.
@yajukun Glad you were able to fix it. Did you try disabling the Nodes 2.0 from ComfyUI? It's in the settings, for me it's the very first option in the settings.
@zkyt Hey, I don't use Nodes 2.0. I think the problem is actually like 2 custom nodes conflicting, and one just happens to be powerloaderv2. Because I have used it previously and I didn't have this issue. But since then I had installed all these other custom nodes for Krea Inpaint, Qwen 2.1, LTX, Lipsync, etc. I really need to consolidate down and just use a handful of models or split up my Comfy installs into groups. Also, I am getting great results with this Ref workflow. Do you think in the future you will add support for refmods? I have a few workflows to create them but haven't tried it yet. Thanks!
@yajukun Ah makes sense. I also just renamed my custom_nodes folder and re-imported only the ones I need. I was actually surprised how fast ComfyUI can boot lol, I totally overloaded my system.
I'm about to test refmods soon, so we'll see.