This series of the models was started (v0.4) from DucHaiten-Pony-Real (beta) by DucHaiten.
Version v1.0 was trained using a LoRa mode in One trainer. Next version are trained in Fine Tuning mode (full SDXL training).
Up to v1.0: Photorealism, lighting and anatomy related LoRas from CinEro series are integrated into this model in order to amplify cinematic render.
Goal of training was to improve to stability of anatomy and make the picture more realistic and moody / cinematic.
Version 1.1: shifted to fantasy / sci-fi themes with less realistic and more contrast / unstable rendering.
Version 1.2 [SOON]: is v1.1 further development with shift to horror / tentacles theme.
PS: be responsible and don't abuse this unrestricted model.
Description
EARLY ACCESS ALPHA (EPOCH 2)
Realistic model sharing some features with CinEro Pony v1.0 ... v1.2
Goal: creativeness + realism + anatomy
PS: it was not trained on xenomorphs...
FAQ
Comments (9)
prompt adherence on 1.3 seems to have decreased significantly. only about 20% of the images seem to adhere. Also noticed more deformity. Still a HUGE fan of 1.1!
Can you give some examples? For me sometimes it shows better adherence for important parts of the prompt.
On the other hand, for the prompt like "Yoda - Hulk hybrid" adherence was lower that I expected.
Honestly, I don't expect much adherence from relatively primitive CLIP in SDXL models.
I prefer to control composition using "Omost" nodes in ComfyUI or using a FLUX generated image.
FLUX encoders adhere prompt spelled in natural language much much better. No chance for Pony to reach that efficiency.
But I am still training Pony CLIP encoders for natural language. Hopefully adherence will be improved.
Thanks for feedback! it is really useful.
https://civitai.com/posts/6686646?returnUrl=%2Fmodels%2F543122%3FmodelVersionId%3D856131
Adherence issue confirmed. In the post above there are 4 imgs from Ep2 and 4 imgs produced by v1.3 Epoch 4. It shows noticeable adherence improvement.
https://civitai.com/posts/6687432?returnUrl=%2Fmodels%2F543122%3FmodelVersionId%3D856131
Look at images with "EPOCH 4" in prompt. For me it seams like natural language adherence increases.
@kaisersose785 I suspect that adherence issue is caused by the bias of the v1.3 toward a human anatomy.
Can you confirm or disproof that?
https://civitai.com/posts/7021057?returnUrl=%2Fmodels%2F543122%3FmodelVersionId%3D887127
I am thinking about merging v1.3 trained with some SDXL universal models (for INxx blocks) to try to fix CLIP issues. May be you can suggest the model (Pony/SDXL) with the best adherence.
@homoludens I agree with your hypothesis. As to your question regarding another model, while underrated, I find the oneTrickPony model to be the perfect balance of prompt adherence and realism.
@kaisersose785 You mean this one?
https://civitai.com/models/648674/one-trick-ponyxl
@homoludens Yep, thats the one