|
--- |
|
pipeline_tag: text-to-image |
|
--- |
|
# wdv-tests |
|
|
|
alright since i've been promising everyone something for ages and i don't have a clue when wdV actual will get done other than soon, here's something that's better than nothing |
|
|
|
## **NOTE:** this is NOT a release, per se. it's a public test, if anything. this uses the TEs from Animagine 3.1 (with permission) and is not representative of the final model really. But it's something, and i rather like it. |
|
|
|
### see below for setup instructions |
|
|
|
- you *need* to use "simple" or "ddim_uniform" or *maybe* "karras" schedule for the sampler |
|
- dpmpp_2s_ancestral works best. **SDE samplers will not work.** |
|
- 20 steps or so at 5.5-7.5 CFG |
|
|
|
## Setup instructions |
|
|
|
### Option 1: SwarmUI extension (recommended) |
|
1. the SwarmUI extension repo is [here on GitHub](https://github.com/waifu-diffusion/SwarmWaifuDiffusionV) |
|
2. Open that link above and read the instructions in the README |
|
3. If you made it to this step, go back to step 2. |
|
|
|
### Option 1: Raw Comfy node (not recommended) |
|
1. get the node from [the WaifuNodes folder of the SwarmUI extension](https://github.com/waifu-diffusion/SwarmWaifuDiffusionV/blob/main/WaifuNodes/wdv_test_nodes.py) |
|
2. put it between your model loader and sampler, |
|
3. set sampling mode to "EDM" |
|
4. set schedule to "XLTC", |
|
5. set scaling 1.6 (these 3 should all be defaults) |
|
|
|
## Model training info |
|
|
|
v0 is trained for five epochs or so on about 4.8 million assorted anime images and uses the text encoder from Animagine 3.1 (with permission from the creator). |
|
Actual release wdV models will not use this TE, to be quite honest i swapped it in for a test and forgot to remove it so here we are. |
