Tired of struggling with workflows trying to get good results with nsfw furries? I've done it for you. Enjoy workflows that will squeeze every last drop of juice out of your computer for incredible quality, with examples on each model (more to come).
Tactical potato for launch that: RTX 4070 with DDR4 3600mhz 96gb ram+5700x3d For fast data exchange with RAM [that L3 3D cache works, damn it. Everything loads so fast!].
(But 64GB will be enough for you, unless it's a Flux 2 DEV and LTX2, i swear).
Description
Note: If you feel like generation is getting slower, be sure to restart Comfy. Unfortunately, it accumulates cache and logs that can't be purge in comfy with nodes. You need to restart GUI. (Comfy problems...ugh, i don't know how to fix that).
I'm having trouble setting up the initial setup of SageAttention and Triton! I've attached a link to a video that will help you set this up for Comfy. On my hardware, I can generate 864x864 videos for up to 12 seconds at 60 fps. It can be longer, but it's quite slow on RAM alone (am on ddr4).
FAQ
Comments (4)
Great workflows I'm New to Chrome but looks brilliant. Currently using the UnCanny (Photorealism Chroma) checkpoint. However, I have a RTX 4090. I was wondering if this workflow takes a while to generate an image or if it's quite quick as I don't have or use sage attention.
Depends on the resolution and settings. Chroma is more flexible for NSFW and has a lot of training material for this. It's slightly faster than regular Flux, but there's nothing stopping you from tweaking a lot of additional finetune. SageAttention It is not necessary element for Chroma.
(Linked tests with/without SageAttention from reddit) https://www.reddit.com/r/StableDiffusion/comments/1l4360d/sage_attention_and_triton_speed_tests_here_you_go/
RTX 3090, 2.29s/it no sage, no Triton
RTX 3090, 2.16s/it with Sage, no Triton -> 5.7% Improvement
RTX 3090, 1.94s/it no Sage, with Triton -> 15.3% Improvement
RTX 3090, 1.81s/it with Sage and Triton -> 21% Improvement
RTX 3070, 7.19s/it no Sage, no Triton
RTX 3070, 6.90s/it with Sage, no Triton -> 4.1% Improvement
RTX 3070, 6.13s/it no Sage, with Triton -> 14.8% Improvement
RTX 3070, 5.80s/it with Sage and Triton -> 19.4% Improvement
I recently switched from uncanny chroma to zbase. Pros and cons to both but zbase is a lot faster
@undertheblacksun Chroma has a Flux Schnell base. It also has a deep knowledge of NSFW, furry, Deepbooru, and E621 tagging, and is easy to train. Yes, Chroma is slower. But this is because it uses negative promt, which significantly improve quality. In Z-Base CFG 1, negative hints are not needed by default, so the initial speed is higher, as with both Flux 2 Dev and Flux 2 Klein. This doesn't mean they can't be used—they significantly improve quality. However, when enabled, Z-Base becomes even slower than Chroma, while significantly inferior in many aspects, except for very low-res quality generation and the initial flexibility and dynamics of poses. Furthermore, Z-Base is censored for furries, and its lore training is broken and cannot be tagged properly. I'd much rather use Klein instead of Z-Base, since it knows what furry is by default and doesn't censor it, simply replacing parts.
I did some additional testing with Z-Base, noticing some oddities at low scale, where I disabled blocks individually. It turned out that Z-Base knows what furry is when block #4 is disabled, but censors them, preventing me from properly creating a furry-NSFW or even just a simple SFW-furry. This is also related to the other elements it ignores.
It really all depends on your preferences and desires. However, among the many models, Z-Base seemed the most "wooden" in its settings compared to others. But by default, if you only need fastgen good-mediocre SFW images and NSFW humans, then yes, it has different material and parameters that you might not find on other models. Plus, it's easier to set up "out of the box". But I prefer to "dig deeper" then others :)