Note: the merge has a slightly more realistic tendency than turbo v1.1. Not the goal (goal was speed),
Settings
text
Steps: 4 (design point)
6 if you need cleaner hands / fingers
CFG: 1.0 for speed
1.0+ if you don't care about speed
Sampler: euler a + simple (benchmarked)CFG above 1.0 doubles cost — it runs a second forward pass per step. Worth it on a fast card, not on a slow one.
Main file is BF16 if you just want the reference weights.
Note: INT8 / FP8 / GGUF files need a loader that supports them. Tested in ComfyUI — other UIs may not load the quantised formats.
Measured performance
RTX 5060 Ti Performance (4 steps, CFG 1.0)
INT8 ConvRot: 2.95 it/s | 2.32 s | ~2.18 GB (Speed Winner)
INT8 (Plain): 2.97 it/s | 2.39 s | ~2.18 GB
FP8 (e4m3fn): 2.75 it/s | 2.48 s | ~2.18 GB
BF16 (Reference): 2.27 it/s | 2.74 s | ~4.18 GB
FP16: 1.99 it/s | 2.89 s | ~4.18 GB
GGUF Q8: 2.02 it/s | 2.87 s | ~2.22 GB
Q5 Tested (1.45 it/s / 6.93 s) but omitted due to dequantization latency.
GTX 1660 Ti Performance (4 steps, CFG 1.0) Flags: --lowvram --force-fp32 --fp32-vae
INT8 (Plain): 5.3–5.7 s/it | ~24 s (Fastest on Turing)
INT8 ConvRot: 6.0–7.2 s/it | ~29–33 s (Rotation adds overhead)
FP8: 13.0–14.0 s/it | ~56–64 s (No native hardware FP8)
BF16: 13.7–14.0 s/it | ~57–67 s (Reference)
FP16: 13.2–14.0 s/it | ~55–66 s
GGUF Q8: 13.2–14.1 s/it | ~58–73 s (CPU dequant bottleneck)
Key Architecture Takeaway: On Turing GPUs (GTX 1660 Ti), Plain INT8 is ~2.4× faster than FP16/BF16/FP8. ConvRot feature rotation creates math overhead on older hardware without Ada/Blackwell Tensor Cores.
Description
FAQ
Comments (7)
This is awesome! It really does produce a great image in only 4 steps. And, personally, I think the baked-in style is much better than Turbo v1.1.
The only thing I'll note is that, like pretty much all the turbo models/loras, if you go above Anima's trained 1.2MP resolution, the image can start to get a little grainy. You need to add a couple steps to compensate. But 6 steps is still a lot better than 16 steps.
Thanks so much for testing it and sharing the feedback!
Honestly, my main goal was just to get this running fast on my old GTX 1660 Ti without dynamic VRAM lag, while tuning it to lock in details at 4 steps. I was actually a bit worried the LoRA bake might push it too far toward realism or trim away some of the original Turbo's style flexibility, so I’m really glad to hear the look lands well for you!
@terpentinas I'll try out some loras and artist tags to see how flexible it is. But, regardless, it's a good look on its own.
@terpentinas Alright, I need to rein in my enthusiasm a bit. I just posted a set of 3 images. Same prompt. Same sampler. No loras. Just different seeds. And I got wildly different styles. The model swings from 2D to 2.5D to near photorealistic. Not sure what happened, but it can't produce a consistent style at all.
@pj1983a That makes total sense! The underlying base model was trained heavily on natural language captions, so raw prompts without strong style anchors naturally drift toward realistic lighting.
Because I baked in detail and de-blur LoRAs, the model stays sharp at 4 steps without strictly needing standard quality tags like masterpiece or best quality. However, those tags also carry a subtle 2D/anime bias. When they're left out, the model defaults straight to that hybrid/semi-real space.
If you want to pull it back toward a specific look:
For 2D / Anime: Adding anime style, illustration, cel shaded or masterpiece helps anchor the look.
For 2.5D / Stylized: Adding 2.5D, digital painting locks in that semi-real look across seeds!
Appreciate you running these tests—this kind of feedback is super helpful!
@terpentinas Good to know!
great model, awesome prompt compliance, well done!