Great job, these models are really an improvement
I have been using the b20-49 version and it works really well, having great image quality while it is able to reference media well. Thank you!
The B20-49 and B25-49 variants are so good that I ended up archiving the default fl2va and ref2va models after running several tests. I don't think it's hyperbole to say that this repo is currently the best thing that the HF community has produced so far for MMH3.
Aside from excelling at standard video generation, it also works extremely well for weird edge cases like image editing in comfyui (using Mamad8's MiniMax-H3-Image-VAE, altering the MMH3 node to allow single frame gen, and setting the output to PNG).
@smhfacct @dampfinchen @fizzlepoof
Can these models be used to feed conditioning with first frame + references? The H3 Comfy nodes don't have slots for both, and the Hybrid Conditioning nodes I've been able to find, seem to break prompt adherence.
@smhfacct @dampfinchen @fizzlepoof
Can these models be used to feed conditioning with first frame + references? The H3 Comfy nodes don't have slots for both, and the Hybrid Conditioning nodes I've been able to find, seem to break prompt adherence.
The models are definitely capable of it. However, if you want to specify a first frame as well as use references (without a hybrid node), then you'll need to use the RF2VA node and then be a bit more verbose in your prompt. Basically, you just need to override the influence of the random seed and other references on the starting frame.
One of the things it likes to do is move the initial position of the camera. Granted, it does this surprisingly well but that's, of course, not what you want for first frame + references. The more references you use, the more verbose you need to be.
As an example:
In the images below, it initially took my input image and moved the camera to a different starting position due to the random seed. When I was a bit more specific, it accepted the input image as the first frame.
@fizzlepoof
I appreciate your input. I haven't been able to get your 1-to-1 reproduction of the input image by prompting Ref2VA.
When swapping the standard Ref2VA model with minimax_h3_hybrid_fl2va_ref2va_b20-49-int8, prompt adherence is destroyed. It ignores my H3-formatted prompt, creating motion & dialogue that are random nonsense.
The standard Ref2VA models do a better job for me, but the start image is always distorted/blotchy enough to be unusable.
I haven't been able to get your 1-to-1 reproduction of the input image by prompting Ref2VA.
...with minimax_h3_hybrid_fl2va_ref2va_b20-49-int8, prompt adherence is destroyed.
...Ref2VA models do a better job for me, but the start image is always distorted/blotchy
There's quite a few things that might cause those issues (i.e. LoRA, heretical CLIP, resolution, sampler, steps, upscaling or lack thereof, etc).
The only LoRA I use is this one. I typically use 25-40 steps, 0.4 to 1mp res, a Spectrum Apply node, a convrot native math CLIP and res_ms with bong tangent or sgm depending on the context. I also sometimes use an ersgan or a latent upscaler if the task requires it. I do use a metric buttload of homemade custom nodes but most of those are QoL or LLM related.
Good luck! 🫡
@fizzlepoof Thanks! 🫡. You use the Kijai Experimental Lora in conjunction with this hybrid model?
The only Lora I'm using is the fl2v_lightx2v_turbo_4step_v0.1, and running it at 8-steps, with everything else at defaults. No upscaler.
EDIT: In case this helps anyone, my distortion issues were from the nvfp4_awq text encoder which is Nvidia Blackwell Specific. I should have been using qwen3vl_32b_minimax_h3_int8_convrot. So @fizzlepoof was correct to mention a CLIP issue (though not a heretical one).