Illustrious-XL SmoothFT
UPDATE: Added SOLID version, a merge of SPO and NoobAI XL v-pred (the resulting model is normal) - it features richer solid colours and smoother gradients.
UPDATE: Added SPO version, thanks to qelt for the suggestion!
About SPO version: This version generally produces better compositions, details and anatomy than the normal version, so I recommend it. The normal version is most true to the original Illustrious model.
Smooth Fine-Tune of the Illustrious-XL base model
This is a personal fine-tune of Illustrious-XL - v0.1 | Stable Diffusion XL Checkpoint
It was trained for one epoch on the same dataset I used for Bijutsubu - v1.0
The fine-tune was then re-merged with the original model using the weights:
0.5,0.3,0.3,0.3,0.15,0.15,0.15,0.15,0.15,0.15,1,0,0,0,0.15,0.15,0.15,0.15,0.15,0.15This was done to retain the characteristics of the original model, while helping fix the block artifact issues I ran into with Illustrious, which are exaggerated by image compression even further:
To illustrate (contrast has been enhanced):
Left: original; Right: SmoothFT
Another example:
The model obviously underwent other changes as a result of the FT, but largely remains true to the original base model.
Thank you to Angel and OnomaAI for this amazing model.
Description
FAQ
Comments (15)
The Solid version is pretty good, thanks! Any chance for a lighting version?
I guess I'll have fun with this SOLID version first before your AnimagineXL+IllustriousXL merge :D
smoothFT-SPOのbasemodelがSDXL1.0のままのせいでgenerationに使うことができません！修正をお願いいたします！
Latest version is very impressive, nice work! But isn't there a way to keep v-pred?
Illustrious-XL SmoothFT
使ってみましたが素晴らしいですね！互換性もバッチリでクリアな絵が出ます！
The 3rd version is ideal for training LoRA! The quality of the final lora + fast training without huge losses during training.
The outfit, appearance and style are perfectly conveyed, if you do it, for example, only on anime screenshots, but if you take only art from artists, then there will be delays with the style, but this is logical, since the art has a different style of artists.
At the same time, if you take art, then everything else is ideal, all the details of clothing and appearance of the characters are conveyed.
Result: I recommend it, since I myself use it as a basic model
Please make the solid version for onsite generation
I just have to repeat, the SOLID model is AMAZING! Please if you manage to keep V-pred somehow it would be the perfect model!
Also, any chance to do a version with the new vpred models? Like 0.65S
bro have you retagged your smooth-ft pro version to Illustrious?, I see it as SDXL 1.0 Illustrious category has been fixed, no reason to keep it SDXL 1.0
Does it require a specific VAE? My outputs get super saturated whenever I try with no VAE or the default SDXL one.
do embeddings made for other models still work? Or would we need ones specifically made for illustrious?
Is it possible to have the "SOLID" version in the site generator?
hi, I use your SOLID version to train my own lora using kohya, but I dont know why my results are not good, https://imgur.com/a/R2UDpDP the illustrious image is far from the original one
hey, is it possible for you to make a checkpoint with the full dragon ball franchise?
A recent update has made this checkpoint an unstable resource. TT
Details
Files
illustriousXL_smoothftSOLID.safetensors
Mirrors
illustriousXL_smoothftSOLID.safetensors
illustriousXL_smoothftSOLID.safetensors
illustriousXL_smoothftSOLID.safetensors
illustriousXL_smoothftSOLID.safetensors
earthboundMagnum_v10.safetensors
illustriousXL_smoothftSOLID.safetensors
illustriousXL_smoothftSOLID.safetensors
illustriousXL_smoothftSOLID.safetensors
Available On (1 platform)
Same model published on other platforms. May have additional downloads or version variants.