Manticore is a custom merge I produced to more accurately recreate traditional styles such as watercolors, sketches, and oil painting.
This model functions very similarly the base NoobAI XL V-pred model, so please follow this guide to get started.
All loras and tags that function on Noob/Illustrious should function just fine.
This model will produce beautiful images of both furry and anime content.
I would personally recommend Euler A sampler with 24-32 steps + normal / SGM uniform scheduler
Enjoy.
Description
Pentacle AKA v5 aims to clean up some of the mesiness or bluriness of v4 and provide crisper, cleaner art that maintains great prompt adherence and style flexibility.
Compatible with NoobAI and Illustrious loras and highly capable in both anime and furry styles.
FAQ
Comments (16)
v5 wahooo!!
Hey, I really like the v4 version,still not trying the new one. Did this model continue adding more furry artists from E621 to the original artists in Noob Vpred? I noticed some artist tags don't seem to be working very well.
unfortunately the training data is the training data. As far as artists are concerned you either hope they had enough images when they pulled the archive like 2 years ago or you use a lora / make the lora yourself.
According to NoobAI XL so long as the artist had about 200 images at the time of training they should "work okay" but there are definitely many that really don't look right or have prompt bleed. Its just an unfortunate reality with the base model itself.
Love v5 (Pentacle) so far! Thank you!
I can't wrap my head why am I getting so different results in Invoke vs ComfyUI (both v-pred turned on). I don't use V-pred much, but Illustrious models give ~99% same results between two apps.
I would really appreciate if you have any tips on this:
Invoke (love the overall dreamy look, lacks black, maybe oversaturated):
https://civitai.com/images/104639645
ComfyUI (sharp, good texture, lacks background)
https://civitai.com/images/104645868
thanks so much!
Do you have a VAE? I didn't see one in the comfy setup, and I've never used Invoke
@Anomalous_Anonymous hmm, just tried "sdxl_vae" and the one found in comfyui workflow NoobAI article (madebyollin/sdxl-vae-fp16-fix) - getting completely the same results in Comfy (and discolored results in Invoke).
I think the comfy workflow is using the same VAE from the checkpoint.
Now I think the comfy results are actually correct, and Invoke handles v-pred differently (tho I liked the results slightly more, and never noticed such difference with other v-pred models).
Thank you! I'll see if I can find out that is the cause of the difference
@Sluggish_Python If its a lack of background on comfy you may be missing the "CFG rescale" node that NoobAI recommends. Invoke probably has it bundled in
@Anomalous_Anonymous thank you! will check what CFG rescale does, never tried that!
I actually just figured out what it was - I had CLIP Skip: 2 in Comfy, and it removed the background (or just didn't recognize "greenery" as a full background) and didn't apply some lighting effects. Which is probably good for the traditional media style :)
I don't use CLIP Skip in Invoke (it's pain to set up), but also don't usually see that much difference if I set clip skip 2 in the same model
Hmm no... I can't reproduce CLIP skip difference anymore with different prompts, this is confusing :D
Probably the combination of a long clunky prompt + CLIP skip 2 sometimes makes a difference
@Sluggish_Python fwiw Invoke does also have CFG rescale if you haven't given that a shot. I believe you can configure it per gen, but I often go to the model organizer and set a default of 0.6-0.7 for any VPred models.
@Kaligo731 thank you for bringing this up! Just tried and CFG rescale does seem to give sharper and more balanced results with this model when using traditional media tags (and little to no quality tags that kill the styling).
Oh and I mentioned here about CLIP Skip in Invoke - found a good explanation in their discord that they use CLIP Skip: 2 for all SDXL models automatically
@Sluggish_Python Glad to hear it's working for you!
I've been testing versions 4.0 and 3.0. Version 4.0 is superior to the original Noobai V-Pred 1.0! It's better quality and anatomy control, so it's become my go-to for everything! Well done!
Can you upload a BF16 version of Pentacle?
I love the style of that dragon girl showcased on the front page here. But running it on Civitai doesn't seem to match at all.
Was there something else you did there?
I was able to get similar images when running them locally in Forge. I believe the Civit generator doesn't handle prompt editing correctly (things of the format [start:end:weight], note the square brackets) which that image uses liberally in its prompt.
Great prompt comprehension, hope this stays up