Snakelite is an offshoot of Snakebite 2.4 that strives for realistic, pleasantly-imperfect photography. It is more experimental than Snakebite and can change drastically from version to version.
Note: All demo images generated using DMD2 LoRA at full strength.
❤️ If you enjoy Snakelite, you can help offset the cost of training:
⚠️ IMPORTANT:
This model uses Flow Matching, so you must connect it to the ModelSamplingSD3 node in ComfyUI to get correct results. Even better is ModelSamplingSD3Advanced.
🤓 Technical details
Over the course of Snakebite 2, I developed a complex merging chain that consisted of around 15 LoRAs - each contributing to model consistency, fine detail, and vivid colors. I'm very happy with the results and I believe v2.4 represents the peak of SDXL realism in some ways.
However, these efforts came at a cost. Snakebite has a tendency to make things look "too perfect" in a Hollywood sort of way. This limits facial diversity and affects compatibility with your LoRAs. It can also suffer from "stylistic collapse": if your prompt is very complex or unusual, Snakebite may struggle to maintain the intended style in its pursuit of aesthetic perfection.
Enter Snakelite:
I have radically simplified my merging chain down to 5 LoRAs. I'm still trying to improve the baseline quality of bigASP, but only in ways that avoid the side effects described above.
I mixed in some Lustify to impart an "amateur photography" feeling. As a result, Snakelite may have potential as a successor to the beloved-but-aging BigLust 1.6/1.7.
Snakelite is still finetuned on the same 1400 images as Snakebite 2.3 for improved aesthetics.
Which version is best is a matter of taste, but I'm guessing there are many Civitai users who may prefer the look of this new model. Let me know what you think!
👍 Advantages over Snakebite v2.4
Colors are more natural, and photographs look more like photographs.
Small faces have more detail.
Better at retaining photographic style even if pushed hard with crazy prompts.
Highly compatible with LoRAs trained on BigLust. (edit - this specifically applies to v1.0, not v1.1)
A bit more tolerant of "quality-boosting" embeddings and LoRAs. You can use DMD2 at full strength or experiment with your own acceleration stacks.
👎 Disadvantages
Not as vivid, sharp, or as contrasty as Snakebite 2.4.
Lower success rate with limbs, extremities, and complex interactions.
Concepts that are unique to bigASP 2.5 are not represented quite as well.
🛠️ Recommended Settings
>=v1.1 (special because they use bigASP 2.6):
8-9 steps
LCM sampler
Beta scheduler
CFG 1
DMD2 LoRA at full strength
ModelSamplingSD3Advanced; decay from 3 to 1
👉 Optimized Workflow (HIGHLY RECOMMENDED) 👈
https://pastebin.com/pZ022SD4
v1.0 with DMD2:
6-9 steps
LCM sampler
Beta, normal, or simple scheduler
CFG 1
Model shift of 3
v1.0 without DMD2:
25-40 steps
Euler ancestral sampler for speed, dpmpp_2s_ancestral for quality
Simple scheduler
CFG 4-6
Model shift of 3
Negative prompt strongly recommended (e.g.
worst quality)
Note: increasing the model shift may improve prompt adherence at the cost of quality. This is particularly useful with character LoRAs. Try a value between 6-8.
Thank you. As always, I look forward to your feedback. Please share the model and upload some images to help it gain traction. It would be amazing if we could make Snakebite eligible for Civitai's onsite generator someday!
Description
Experimental update that replaces bigASP 2.5 with a preview version of bigASP 2.6. To my knowledge, this is the first checkpoint on Civitai based on 2.6. 🫡
First impressions: the preview checkpoint may be a little overtrained... it's great at sharp, vibrant photography but it can be difficult to prompt for subtle/soft textures like you can with Snakelite 1.0. That said, overall prompt adherence and backgrounds are phenomenal for SDXL.
Also, if you require compatibility with your old LoRAs - stick with 1.0. The new version produces very different results.
Check model details for tips and new workflows. Thank you.
FAQ
Comments (5)
Sometimes the images come out overcast or sepia-like, do you know of a good way to prevent the more realistic or natural look?
Also, any tips for inpainting?
Snakelite is also really neat btw
Hi XpomulCivi, I know what you mean about the "overcast" look and I'm still experimenting with different merging strategies to see what works best with the new bigASP 2.6. In the meantime, you can try the Offset LoRA which behaves as a sort of brightness/exposure slider.
Increasing your CFG will also, in general, make images look less realistic.
As far as inpainting, one trick that works really well is the Multiply Sigmas (Stateless) node from ComfyUI-Detail-Daemon:
- factor = 0
- start = 0
- end = 0.4 (lower for "more creative," higher for "more like the starting image.")
I would also suggest not using the included VAE for inpainting, as it tends to make images look green and a bit glitchy. It's only really good at txt2img. So you'll want to load a separate VAE like AiArtLab/sdxl_vae.
Hope that helps!
@liftweights it does, thank you again!
This is a really intriguing checkpoint, I was thinking of trying to to train loras on it.. but alas it seems to not work. Is this due to the flow matching? Is it not possible or am I just doing something wrong? I tried using the base; snakebite, but yeah images just came out as blob and color, not undertrained, but like it just didn't know what to do with it. lol
Hi @Light7799 , your hunch is correct: the training approach must support flow matching. Check the bottom of Snakebite 2's model description for recommendations. I've trained a few character LoRAs on v2.4 and they all turned out really well, but it does take a little elbow grease to set up the environment.
edit - It's also important to train on a non-turbo version of Snakebite. Either Snakebite 2.4 Full or Snakelite 1.1 should work fine.