CLIP (ViT-L/14) is a text encoder that all of these stable diffusion models were trained on. All of SD 1.0 through 1.5 were trained on a fixed CLIP. LoRAs, Finetunes, and merges on the other hand modify CLIP to add new concepts. The result? In our model universe, our CLIP has been pretty abused and often strayed too far from the original. I initially merged a lot of CLIPs I thought worked well for certain things, but now I realize that we have to start feeding v1.5's CLIP back in to move forward.v1.7 partially merges in bits from three of my experimental models:heater, a custom model sampled from 47 anime models; dare-ties at 8%.fullpoint, a custom model sampled from 37 other general models I had fp32 data for; dare-ties at 8%..darkness, a custom model sampled from 57 artistic models; block merged in [30%, .20%, 10%].Really just a taste of each. Since all of these models are so strongly pooled, adding each one in mixed in some really strong core structure. If you look at the workflows in the images, you can see the CFG is pretty high, as are the number of steps (50 seemed good for 768x1152), using dpmpp 2/3m sde and CFG of 10+ in parts. I even got decent images using exponential sampling in some cases. Artifacts on faces are very rare, and hands generally work pretty well.I also 50/50 merged my cybernetic dreams CLIP, cleaning up a lot of strange issues.That structure I added should make this version much more stable when using both LoRAs and LCM as well.v1.6 attempted to improve hands and other details. It uses the new DARE-TIES merge to cherry pick these while maintaining some of the stylistic qualities; along with some creative use of the skin-hands LORA (<a target="_blank" rel="ugc" href="https://huggingface.co/imagepipeline/Skin-and-Hands">https://huggingface.co/imagepipeline/Skin-and-Hands</a>), which is very good. I've been using <a target="_blank" rel="ugc" href="https://civarchive.com/models/116230?modelVersionId=125849">https://civarchive.com/models/116230?modelVersionId=125849</a> bad hands 5 to clean up most of the remaining glitches.v1.5 applies <a target="_blank" rel="ugc" href="https://huggingface.co/bdsqlsz/dpo-sd-text2image-v1-fp16">https://huggingface.co/bdsqlsz/dpo-sd-text2image-v1-fp16</a> DPO model to provide some more structure.This model's lineage is originally DarkSun, as I liked it's style, but it may have some very different powers as it has been subjected to a lot of theories that I have about how diffusion models work (and a lot of merging)... v1.4 was an experiment using FreeU as a model preprocessor, but upon further testing, it doesn't seem to have an impact on the final results. Why? FreeU is an outblock preprocessor.For any users, post your best images (with prompts) and I will use them for DPO in the future.Hearts/5-stars if you want me to continue working on this one.Try different clip skips, can handle very high CFG (9+) and really aggressive sampling.Disclaimer: This model is not intended to produce CSAM. Do not use it for this purpose.

DarkClip 2.5d

The model was missing a key (text_model.embeddings.position_ids), which comfy currently drops and automatic expects, so I patched it back in. I also changed the bundled vae to vae-ft-ema-560000-ema-pruned.

Files with SHA256: 754e18ba3f55221a778e3714b4469eebc8a9ee73c2acaf6a5f239ba81396adda