This is NOT a basic workflow.

This is NOT meant as a low VRAM workflow.

This is NOT meant to be a workflow centered around specific base models or LoRAs and is not meant to be a prompt and click generator. I experiment a lot and this workflow is meant for this purpose.

There are tips and instructions in the workflow, so I won't elaborate much on instructions for use. I don't know if this is the right way, but it's my way. :)

This Workflow includes:

1. Custom Node: H3 Keyframes (keep TE) is a wrapper for the H3 Keyframes node from ComfyUI-H3-Multishot, built for multi-GPU setups.

   The stock node moves the text encoder to system RAM after encoding and purges ~90% of VRAM on the main GPU. On a single GPU that's helpful, but with the text encoder on a second GPU (MultiGPU/DisTorch), it pulls the encoder off its dedicated card and can unload the video model, forcing both to reload every run.

   This node adds two toggles:

   - keep_text_encoder (default on): leaves the text encoder on its GPU.

   - vram_cleanup (default off): skips the VRAM purge.

   It calls the original node rather than modifying it, so pack updates still carry through. Requires ComfyUI-H3-Multishot. Unzip the H3-Keyframes-KeepTE.zip directory into your custom_nodes directory.

2. LoraManager Lora Loader and TriggerWord Toggle, because if you aren't using LoRA Manager, WTF are you doing?

3. MultiGPU Model Loaders

4. Sage Attention Options:

• MiniMax Mem Eff Sage Attention

• Patch Sage Attention KJ

• Comfy Kitchen Attention

1. MiniMax H3 Chunk FeedForward and MiniMax H3 Low VRAM Attention

2. Impact Wildcard Processor for wildcard {|} prompts

3. DaSiWa Enhanced Video Combine because it is really good.

4. Model Preview Override

5. Start Image input, 6 reference image inputs and 3 Audio inputs

I don't have an Upscaler included because I use a separate workflow for upcaling and interpolation. https://civarchive.com/models/2972841/video-upscale-and-interpolation

Primary groups are free floating to arrange wherever you want without breaking the workflow and no annoying spline webs.

There are links to basic base, text encoder and VAE models as well as Accelerator LoRAs in the workflow. I am leaving it up to you to find the non-standard models if you want to use them and ComfyUi makes that pretty easy now.

Please make sure you look at the instructions for the models you want to use for best practices for steps, scheduler/sampler and cfg settings.

if you want instruction for any of the nodes I use, please look to the creator's documentation. I do have suggested settings for different model profiles in the workflow note nodes. Example images have metadata and workflows embedded.

This workflow is a work in progress and I will update it with any major changes as they happen.

Feel free to message me and ask questions and I will do my best to help you when I am able.