---

license: apache-2.0 base_model: Qwen/Qwen3.5-2B tags: - layout-planning - prompt-upsampling - distillation - bbox

pocket-layout-planner-2B

A Qwen3.5-2B fine-tune that takes only a short image request (no system prompt, no thinking) and returns, on one line, a detailed caption citing elements as <id> followed by a JSON box table placing each element — [y_min, x_min, y_max, x_max] on a 0–1000 grid, top-left origin. Distilled from the t2i_bbox mode of black-forest-labs/flux-3-image-klein-pup (Apache-2.0, credited as teacher).

Evaluation (300-request stratified hold-out + 100-prompt image comparison)

Text-level — all 301 hold-out requests, greedy:

Metric	2B SFT (this model)	0.8B SFT	2B untuned + t2i_bbox prompt
parse + fully-cited rate	92.7%	95.0%	1.0%
valid-box rate	95.7%	96.4%	10.6%
quoted text verbatim	55/55	55/55	55/55
mean matched-row IoU to teacher	0.529	0.470	0.133
mean output tokens	646	618	810

Image-level — 100-prompt comparison (60 GenEval-subset + 40 spec requests), Klein fp8r 512², 1 seed, "GenEval-style (OWLv2)" re-implementation (not official GenEval):

Arm	GenEval-style (60)	Spec-request object counts (40)
raw request	58/60	29/40
pup plain t2i	60/60	30/40
pup t2i_bbox (teacher)	60/60	28/40
0.8B SFT	53/60	26/40
this 2B	58/60	26/40

Finding, stated plainly: no layout advantage materialized — for the teacher or the students. These models are best understood as pocket prompt upsamplers that emit a structured, parseable layout record (96%+ reliable), not as layout planners that improve adherence. The planned DPO stage was dropped as moot. Contact sheet and per-prompt data: ysharma/klein-layout-plans.

Training recipe

• Identical to the 0.8B except: full fine-tune on an A100 (1h1m, 3.68 s/step), 2 epochs (10.1M tokens), same 7,575 filtered rows, LR 1e-5 cosine. Final train loss 0.724 (token accuracy 0.810); checkpoint-469 is the epoch-1 fallback.

Executed snippet

See the 0.8B card — identical usage with model_id = "ysharma/pocket-layout-planner-2B".

Limitations

• Layout does not improve Klein distilled renders — use as a compact upsampler, not an adherence-boosting planner. Full-size/base Klein not measured.
• Object-count satisfaction of renders trails the raw request slightly (26/40 vs 29/40).
• Output is English; requests in any language.
• The image comparison is 100 prompts, 1 seed, presence-count scoring — far below the official 553 × 4-seed GenEval protocol.

Licence

Apache-2.0 (inherited from Qwen3.5-2B). Teacher credit: black-forest-labs/flux-3-image-klein-pup (Apache-2.0). Not affiliated with or endorsed by Black Forest Labs.