Has anyone else tried RefMod. I think it's very good. https://youtu.be/2K6-OtV_Vbc If you could incorporate it into your pipeline Andy, I think you would have a powerful option and feature to offer. It's like a quick Lora, you dump all the reference images into a folder, any size, no description or naming convention. Point the RefMod workflow to the folder, and in moments you have a .safetensor file you can load into a RefMod loader node and then into your workflow. I got ChatGPT to customise the H3 Director (seesee75 version) node to add the RefMod loader output (or multiples) as inputs to the Director node, just like any other ref image input. It's like having 20+ reference images per character, location or prop in one input. I have the custom minimax_director.py (replacement) file if anyone wants to experiment with it, but you can also test it in its native workflow. * you can also add audio and video ref samples to the RefMod creation.

This is a 3+ minute music video. Pretty seemless (a little more accurate prompting needed). The only issue now is that on these long renders the face starts to degrade. Take a look.

Hi, I raised the issue of the candidate Refine step on your YouTube video comments. How can I provide you with more info to figure out the fault? Thanks

Minimax H3 Director v1.4 second pass workflow issue Hi Andy - I'm using the prompt override box for test prompts (Prompt override from socket is checked) - without the second pass, the preview and final output all work ok. I'm setting 18 steps total; 14 steps for initial pass, then 4 for second refine stage. If I check Enable Two Stage Sampling, it crashes after generating the 14 steps first pass (no video output). I'm guessing it should do 4 more passes, then move to the latent upscaler? But the run crashes out; seems to be on the temporal chunks checkbox. Doesn't matter whether the temporal chunking is checked or not, same crash. Suggest the second refinement pass and the upscale be separated? (I might have it wrong, thought refinement passes were using the same first-pass model? Or is it trying to use the upscaler model to do the refinement and upscale at the same time?) [INFO] got prompt [INFO] [MuseMinimaxDirector] mode=Reference (Omni) — up to 7 images, 3 videos, 3 audio, 1376x768, seed=1111429277178777, seed_hunt=False, candidates=1, 1 chunk(s): ~15.0s (total 15.0s) [INFO] [MuseMinimaxDirector] chunk 1/1, seed=1111429277178777, length=362 frames, video_carry=False, audio_carry=False, hybrid=False [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] Model MiniMaxH3VideoVAE prepared for dynamic VRAM loading. 4965MB Staged. 0 patches attached. Force pre-loaded 128 weights: 348 KB. [INFO] Model MiniMaxH3VideoVAE prepared for dynamic VRAM loading. 4965MB Staged. 0 patches attached. Force pre-loaded 128 weights: 348 KB. [INFO] Model MiniMaxH3VideoVAE prepared for dynamic VRAM loading. 4965MB Staged. 0 patches attached. Force pre-loaded 128 weights: 348 KB. [INFO] Model MiniMaxH3VideoVAE prepared for dynamic VRAM loading. 4965MB Staged. 0 patches attached. Force pre-loaded 128 weights: 348 KB. [INFO] Requested to load MiniMaxH3AudioVAE

Building the roadmap around what's actually useful to you. Pick the one you'd get the most value from next, and I'll try to turn it into tutorials/workflows.