Minimax H3 advanced workflow for creating long videos with motion context (Nikodemon80's H3 Motion context custom node) using another generated video or starting from scratch. No need to save and load latents at each stage. Generation can be resumed from any stage if any changes are needed using ComfyUI's internal mechanism.
Basic functionality:
1. Choose between Sampling or "Start with video" depending on if you want to start from scratch or from a generated video.
2. If Sampling from scratch enter width/height in nearby blue widgets and images in first/last frame. In "start with video" mode, they are bypassed and taken from video size.
3. Enable stage groups for number of generations.
4. 4 step turbo LoRA is applied in models subgraph.
5. Separate LoRAs can be applied to each stage using the power lora loader node in each stage.
6. Generation can be resumed from any stage if you want to change something.
7. Generation can be extended to your device's capabilities by replicating "Stage 4" group any number of times and connecting in similar fashion.
Description
FAQ
Comments (9)
Hi! Is this exactly what I’ve been looking for?
Can this workflow use, for example, a 20-second reference video and replace the person in it with a person from a reference image, while processing the video in smaller sections so that the VRAM doesn’t run out?
In other words, I’d like to recreate the reference video while replacing the person in it.
I haven’t found a single workflow yet that supports this. With 16 GB of VRAM, I can only process around 8–10 seconds from a reference video without manually splitting everything up, which unfortunately makes the final video look choppy.
That is REF2VA, this is FLF2VA. The model has multiple layers. At least a minimum few layers of the model have to fit in VRAM along with video data, so manual splitting is needed. Though you can try by reducing the dimensions of input ref video.
You can try with the REF2VA workflow in example workflows directory within custom node of NikoDemon80's H3 motion context but you might still have to split it.
I posted a face swap workflow, it's on my models page, in the description there is a link to a full character replacement WF. Same method. Just use Meta Batch Manager node if you need to chunk your frames. In general, any WF. Connect it it to VHS loader and combine nodes, set the number of frames to process in a batch. It handles the rest. As long as you adhere to H3 frame count rules. If you have and ending chunk that doesn't align, pad some duplicate frames on the end, slice -1 maybe, then repeat image batch with an expression like (17 - ((a - 5) % 17)) % 17. If a is the MBM batch size, that formula should spit out the exact number of repeat frames you need to pad to adhere to the rules and not throw an error.
Hi, Thanks for the WF, quick question, I'm getting an error of "Context Frames" missing between stage 1 and stage 2. Can you help? All the nodes seems conected. No changes made to your default WF. Thanks for reading this.
If the error is regarding "frame_count", make sure you are using the latest version of NikoDemon80's custom node. The workflow has been tested before upload.
Thank you very nicely assembled
I really like the look of this, hoping I can get it going! Question, do my first and last frames need to be the same size as the video or will it resize them for me?
Great! I really like it – I just added some CleanVRAM because I was experiencing quite a bit of slowdown in Videocombine, but everything’s running smoothly now. Thanks.
I get motion context errors