ysharma/agate-002-4step

Agate 002 — 4-step distilled (unofficial)

An unofficial distillation of LogoLabs/agate-preview-002 (MIT) into a 4-step uniform-grid Euler velocity model with classifier-free guidance baked in (4 network evaluations per image vs the teacher's 100). An 8-step variant is included under 8step/. It drops straight into the Agate webgpu page with steps=4, cfg=1.

from agate import AgatePipeline
pipe = AgatePipeline.from_pretrained("ysharma/agate-002-4step")
images = pipe("a red cube on top of a blue sphere", steps=4, cfg=1.0)   # one image, no CFG

Method: guidance distillation (teacher CFG-3 velocity target, frozen Ettin text encoder, AdamW 1e-5, EMA 0.999) then progressive distillation 32 -> 16 -> 8 -> 4 (Salimans & Ho 2022; Meng et al. 2023), trained on 150k SD-VAE latents from LucasFang/FLUX-Reason-6M (Apache-2.0) with Agate's caption mix (80% short / 20% detail).

Results (256 px, 1 image/prompt, watermark off)

config	steps	net evals	GenEval	CLIP	FID-5k	FD-DINOv2	LPIPS-vs-t50	s/img
teacher50	50	100	0.562	27.03	11.40	130.78	-	0.591
teacher16	16	32	0.569	26.74	11.90	129.98	0.196	0.189
teacher8	8	16	0.533	26.67	13.47	141.18	0.298	0.095
teacher4	4	8	0.466	26.62	20.07	218.23	0.416	0.047
G @50 (GD student)	50	50	0.507	26.39	11.89	131.86	0.218	0.580
s8 (8-step student)	8	8	0.521	26.49	12.35	134.64	0.262	0.093
s4 (4-step student)	4	4	0.527	26.25	12.96	146.47	0.307	0.046

GenEval = official scorer (Mask2Former-Swin-S + OpenCLIP ViT-L/14), mean of the 6 task accuracies; the teacher's published 0.577 uses the official 4-image protocol (0.562 here with 1 image). FID-5k / FD-DINOv2 are against the 5,000 held-out FLUX-Reason images at 256 px and are only comparable within this table. CLIPScore = openai/clip-vit-large-patch14, x100 scale. LPIPS (AlexNet) = vs teacher50 on the first 500 GenEval prompts, same seed. Latency = batch-1 a100-large with CUDA graphs.

At the same 8 network evaluations, the 4-step student beats the teacher run at 4 steps on GenEval (0.527 vs 0.466), FID-5k (12.96 vs 20.07) and FD-DINOv2 (146.5 vs 218.2), and lands closer to the teacher's 50-step output (LPIPS 0.307 vs 0.416); it trails it on CLIPScore by 0.37 points (26.25 vs 26.62).

Limitations

• No negative prompts: CFG is baked in; negative_prompt is ignored at cfg=1.0.
• 256 px only, like the teacher.
• Inherits the teacher's weaknesses (counting above ~3, negation, exact text) and its unprompted-nudity tendency — use an output classifier before public deployment.
• Guidance-distillation gate (LPIPS G@50 vs teacher@50-CFG) measured 0.174, above the 0.15 target; stage LPIPS vs each stage's teacher: s16 0.015, s8 0.032, s4 0.104.
• The watermark (payload AGATE002) is on by default; pass watermark=False for images that go into metrics.

Credits

Model and code: LogoLabs (Logolabs/agate-preview-002, MIT). Unofficial distillation by ysharma; distilled weights under the same MIT license as the teacher's code.

Images

The same 16 prompts and seeds, left to right — teacher (002) at 50 steps with CFG 3, the 8-step student, the 4-step student:

teacher @50 (CFG 3)	s8 @8 steps	s4 @4 steps

Training

Per-stage distillation loss (log scale), MSE against each stage's teacher target:

Stage checkpoints and logs: ysharma/agate-002-distill-artifacts (g/, s16/, s8/, s4/, logs/, grids/). Live dashboards: agate-002-distill-trackio. Evaluation data and raw results: ysharma/agate-002-eval-images.

Stage LPIPS against each stage's teacher (8 fixed prompts, same noise): s16 0.015, s8 0.032, s4 0.104. Guidance-distillation gate: G@50 vs teacher@50-CFG 0.174 (above the 0.15 target — recorded as a finding); G@32 vs G@50 0.004.