Note: the merge has a slightly more realistic tendency than <a target="_blank" rel="ugc" href="https://civarchive.com/models/2458426/anima?modelVersionId=3263843">turbo v1.1</a>. Not the goal (goal was speed),<hr /><h2 id="settings">Settings</h2><span style="color:rgb(249, 250, 251);font-family:Menlo, Monaco, Consolas, &quot;Cascadia Mono&quot;, &quot;Ubuntu Mono&quot;, &quot;DejaVu Sans Mono&quot;, &quot;Liberation Mono&quot;, &quot;JetBrains Mono&quot;, &quot;Fira Code&quot;, Cousine, &quot;Roboto Mono&quot;, &quot;Courier New&quot;, Courier, sans-serif, system-ui;font-size:12px;line-height:18px">text<pre><code>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)</code></pre>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.<hr />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.<hr /><h2 id="measured-performance">Measured performance</h2>RTX 5060 Ti Performance (4 steps, CFG 1.0)<ul><li>INT8 ConvRot: 2.95 it/s | 2.32 s | ~2.18 GB (Speed Winner)</li><li>INT8 (Plain): 2.97 it/s | 2.39 s | ~2.18 GB</li><li>FP8 (e4m3fn): 2.75 it/s | 2.48 s | ~2.18 GB</li><li>BF16 (Reference): 2.27 it/s | 2.74 s | ~4.18 GB</li><li>FP16: 1.99 it/s | 2.89 s | ~4.18 GB</li><li>GGUF Q8: 2.02 it/s | 2.87 s | ~2.22 GB</li><li>Q5 Tested (1.45 it/s / 6.93 s) but omitted due to dequantization latency.</li></ul>GTX 1660 Ti Performance (4 steps, CFG 1.0) Flags: <code>--lowvram --force-fp32 --fp32-vae</code><ul><li>INT8 (Plain): 5.3–5.7 s/it | ~24 s (Fastest on Turing)</li><li>INT8 ConvRot: 6.0–7.2 s/it | ~29–33 s (Rotation adds overhead)</li><li>FP8: 13.0–14.0 s/it | ~56–64 s (No native hardware FP8)</li><li>BF16: 13.7–14.0 s/it | ~57–67 s (Reference)</li><li>FP16: 13.2–14.0 s/it | ~55–66 s</li><li>GGUF Q8: 13.2–14.1 s/it | ~58–73 s (CPU dequant bottleneck)</li></ul>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.

Anima Turbo 4S + Quant Pack

anima-turbo-v1.1-4s-fp16.safetensors

anima-turbo-v1.1-4s-merge-bf16.safetensors

Files with SHA256: 2898bf510e28b7cad364a6815a887e0cf24cb3143e176b89a49dd471f8d39dff