Like some of ya’ll I have been having fun using the H3 model to mess around with so I have been experimenting with using H3 model to be a consistent character generator which leverages multi image reference (up to 9), so I made a workflow which you can use ‘less than ideal’ images from google to build a consistent character and output a 360 character sheet to use as a reference sheet for future H3 generations.
The goal is to achieve high character consistency across future generations. I have tried my best to keep the workflow simple without too many custom nodes.
How it works:
- You input your images and describe them in the Input text section (A Prompt)
- The text is combined with a fixed prompt which spins the character (B Prompt)
- The video is generated at a slow speed with no hard cuts (only camera spin and pan) to maintain character consistency
- Image is assembled with optional character video and full individual frame output (if you want to use for future)
I have included a 6 panel WF and a 4 panel WF. The 4 panel works faster by generating 40% less frames.
Current Caveats:
- The model is quite slooooow. You are also generating 124 frames only to use 6. I have partly solved this by also uploading a 4 panel version.
- Speed ups (like Turbo LORAs) help with speed, but it hurts prompt adherence and quality slightly.
- Quality is limited, since it is a video model it is better at generating video than images. You can solve this by generating at a higher resolution at the tradeoff of longer gen times. You can also use the individually split frames as future references too.
- Details when using this character sheet as output for future generations on H3 may also be limited due to resolution also, I recommend you use this character sheet (for consistency) + other images close up angles (i.e clothing details/face) if doing close ups. If you are just doing a one off video you may possibly be better off not using this character sheet.
I have also included a modified B prompt to do Anime2Real since someone asked for it. Working on tidying it up a bit more.
Link to the 4 and 6 panel workflow can be found here: https://huggingface.co/PoopMan333/H3_Character_Sheet_Generator
Some notes I just remembered:
- You can increase the steps and it may improve your quality slightly.
- With the Turbo Loras enabled, prompt adherence sometimes suffers, but you may be able to get a good seed with another roll of the dice.
- Currently the B prompt specifies a “neutral A pose”, please remove this if you want your character in a particular pose.
- You can use a few different shots of the same character to reinforce the 360 and get more accurate details right.
- Can be used for objects / props also, may require some changes to the B prompt.
