This is a port to the H3 FL2VA model of Wan 2.2 I2V - Multi First-to-Last-Frame (FFLF) video. Automated, Easy, Looping, Multi-Prompting.
It automates the production of any type of long cinematic sequence, such as photoshootings, using the First-to-Last frame technique and a bunch of keyframes.
Settings
Load Images From Folder (KJ)
folder: set the location of your images
load_image_cap: the number of images to run the workflow on.
start_index: skip the first n images.
Looping Video
if set to true, the workflow will do an extra loop to link the last image of the dataset to the first one.
SetImageSize
width and height must be multiples of 32.
Master Prompt
Customize the prompt here. It will be injected at every run.
Specific Promps (1 per line)
Add prompt instructions for specific runs. The first line will be used at the first run, etc.
Lengths (17n+5)
The number of frames to generate. Typically one value of 22, 39 or 56. Can be customized for each run.
Usage Tips
RAM
While the workflow can theorically run on an infinite number of images, hardware limitations will hit at some point. For Windows users: make sure to have set up a huge pagefile.sys file before the run.
Audio
Since each image pair is generated separately, the audio generated by H3 will be inconsistent. Therefore, the audio outputs are disabled, and replaced by a Load Audio node.
Transitions
If they feel unnatural : try changing the order of images, and make sure the various elements, such as the background and the clothing, are consistent from image to image.
If they feel too slow or too fast : try adjusting the chunks' lengths individually.
Turbo
I got the best results using:
minimax_h3_fl2v_lightx2v_turbo_4step_v0.1_comfy lora, at 0.75 strength and 8 steps.
Interpolation
It is not essential since the videos are generated at 24 fps, but interpolation helps at smoothing transitions.
Feel free delete, or disable and disconnect the subgraph if not used.
Misc
Skip 1st/last frame: not sure which one works best, but better keeping one of them enabled.
Save Chunk: records each generated chunk individually.
Preview Whole : gives a preview of the combinated chunks.
Going Further
I believe the generated output, along with the audio and a few keyframes, could be used with the Ref2VA model for further refinements (such as smoothing transitions up and lipsync), but I haven't managed to produce anything meaningful with it to date.
Dependencies
Required
Optional
RES4LYF (to set the output's size)
ComfyUI-GIMM-VFI (for interpolation)