Updating for Wan! All Wan models are compatible with 2.1 and 2.2.

If you want to use Anna, triggers that work are Disney Anna, Disney Princess Anna or CGI Disney Princess Anna.

So it was trained with "some" 2D, more of a stylistic choice for me than a necessity. Prompting, cartoon, line art drawing, disney style animated video will help with that. I may do the 2d style separately when I have more project time available.

She is in there.

https://civitai.com/images/57056416

@Suppressor Thank you a lot :)

Funny you should ask, I am captioning training material for Holly Would right now.

@boxertwin75593, Holli Would is coming tonight. Just got done training, just getting example images ready.

He did a Holli Would: https://civitai.com/models/1336657/holli-would-cool-world-hunyuan?modelVersionId=1509433

You legend!

Thanks I appreciate. I'd love to dive into Pixar girls once I get done with my current styles project.

The trick is heavily captioning for the character's name and like 2 traits that are unique to them. This model can still get some troublesome feature blending if you mix ethnicities for some reason, Like Jasmine and Elsa. But mostly it seems to overcome it with careful prompting.

@Suppressor Alright, thanks! I suppose I have to up my captioning game then

Likely. If WAN convinces me it's better, otherwise I'm not going to waste the time.

@Suppressor I've been testing since release day and it definitely is better.

@azeli Any particular recommended checkpoints or workflows? The only thing I see that is preferable is the slightly longer stable generation times. The quality seems exactly the same to me so far.

@Suppressor It's not the quality per se, it's the movement which is much more natural. There are plenty of comparisons here on and reddit that show even without optimizations and few loras it's still clearly better.

I'm testing all kinds of checkpoints at the moment. The best would be the FP16 just released a few days ago by Comfy, if you have the VRAM for it.

Workflows I would avoid many here and just built your own based off the native example. Teacache works which reduces speed by half if not more.

@azeli Ok. I'll do training and testing on one of my smaller lora projects first, then I'll redo the princesses for Wan if I like the results.

@Suppressor I would be really interested and happy to test with you, DM me if it's easier. But essentially I've found that all current Loras on here ruin the results of I2V.

Try a basic I2V using a high quality Flux image, even on the 480p the results are absolutely stunning quality. Then add a lora and you'll see the drop is horrendous, completely ruining the video.

If you're a seasoned trainer it may be a good challenge to produce a high quality lora, and I believe this isn't something people are aware of or discussing yet.

@azeli So it looks like i2v and t2v are not compatible with lora training. You can either train for one or the other, since you have to use a different model to train each. This is unfortunate. It means you have to train a version for each, double the work. I mostly intend to do t2v, at least at first.

@azeli just started training for Wan2.1, going to try and really hit the quality notes with this config:
output_dir = 'output' dataset = '/eval_dataset.toml' epochs = 75 micro_batch_size_per_gpu = 4 pipeline_stages = 1 gradient_accumulation_steps = 8 gradient_clipping = 1.0 warmup_steps = 200 blocks_to_swap = 10 eval_every_n_epochs = 1 eval_before_first_step = true eval_micro_batch_size_per_gpu = 2 eval_gradient_accumulation_steps = 4 save_every_n_epochs = 3 checkpoint_every_n_minutes = 60 activation_checkpointing = true partition_method = 'parameters' save_dtype = 'bfloat16' caching_batch_size = 4 steps_per_print = 1 video_clip_mode = 'multiple_overlapping' [model] type = 'wan' ckpt_path = 'chkpts' dtype = 'bfloat16' timestep_sample_method = 'logit_normal' [adapter] type = 'lora' rank = 64 dtype = 'bfloat16' [optimizer] type = 'adamw_optimi' lr = 1e-5 betas = [0.9, 0.99] weight_decay = 0.01 eps = 1e-8

Are you 100% sure, because they "kind of work" on both, I've tested. Albeit currently neither gens are that great for the lora i'm testing but I do wonder if it's the lora.

Half my issues I think are lora related right now.

DM me if you want someone to test, I've just got 2 x 5090s

@azeli My understanding is that because they are 2 different base models, and you can only train for one (for now), that the results will be biased in favor of the model you train on. That may change if someone distills them together perhaps?

@azeli Right on with the testing, I will shoot it to you first. I am using A100 SXM to train, but I can only use it at certain times, so testing is done on my 4080.

https://civitai.com/images/56680809 was the best generation I got for her. It really helped to prompt for her outfit.

Also specify ethnicity and eye color specifically

@Suppressor Thanks, I'll give it a shot.

Hi, thank you. I use diffusion-pipe for my training. How many images are you training at 100 epochs?

@Suppressor hi, In the one I tried to make, I used 30 images with good quality, but the clothes didn't match the images and that's even though I manually modified the txt files. Maybe it's because of OneTrainer, although now it doesn't work for me. I suppose it's because of an update that broke it and it always gives me a VRam error even if I set the values ​​almost to minimum because the configuration I had saved and that worked for me doesn't work now. I'll have to try installing Pipe, although I'd say it's somewhat complicated.

@stylobcn VRAM errors are tough to debug, I know. I would bump up your image count a little and caption for the clothes so it knows they are not part of the lora (unless you want them to be.) Pipe is tougher to learn, but a lot more rewarding when you get the hang of it.