Oshikaka/ultraedit-api-judge

UltraEdit outputs for GPT / Gemini success-rate evaluation

This repo holds the edited images and videos behind the result rows of our paper (UltraEdit, the baselines, and our controls), plus scripts that measure edit success with the OpenAI (GPT) and Gemini APIs. Each output is judged on its own: the judge sees the source, the output, the edit instruction and the benchmark's success question. It never sees the method name.

What to test, in order

Every task has a tier and a paper_row (the table and row it supports). Run tier 1 with both providers first, then tier 2, then tier 3. summarize.py groups the results by paper table.

Tier 1: the main tables (--tier 1)

• Tab. 1 (tab:main), SPRITE (491 edits) and PIE-Bench (700 edits): rf_solver, fireflow, flowedit, syncsde, samflow (baselines), ultraedit (ours), ultraedit_box (CLIPSeg box, no return) and flowedit_t1_g35 (a measurement of FlowEdit: started at pure noise, t = 1, with one guidance 3.5 / 3.5; in the paper only in the supplement, tab:fe-curve and tab:pie, and in the text of Supp. supp:design). Claim: UltraEdit makes many more of the requested edits than every baseline.
• Tab. 3 (tab:video), FiVE-Bench (419), NRVBench (180), IVEBench motion (59 x 3 seeds): wan_edit, ultraedit, ultraedit_reanchor (FiVE row; NRVBench in the supplement) and wanedit_t1_g5 (a measurement of Wan-Edit: started at pure noise with guidance 5 for both copies; in the paper only in Supp. tab:supp-wan-t1). Claim: UltraEdit makes more of the requested edits than Wan-Edit, motion edits most of all.
• The key comparison is ultraedit vs flowedit_t1_g35 (and, for videos, vs wanedit_t1_g5). Started at pure noise with one guidance, the other editor makes about as many edits (Qwen judge: SPRITE Edit 0.780 vs 0.784, PIE-Bench VQA 70.0 vs 69.8). At this matched edit success, does the API judge also find that UltraEdit keeps the subject and scene (preserved) and makes fewer artifacts (artifacts)? summarize.py reports this on all edits and on the outputs both methods pass.

Tier 2: Tab. 2, the design space (tab:design, SPRITE only; --tier 2). The design-space rows: ultraedit_whole_forward, ultraedit_whole_fresh, ultraedit_region_fresh, ultraedit_region_annealed, ultraedit_region_fixed, ultraedit_outside_forward, ultraedit_no_return, ultraedit_no_kv, ultraedit_no_kv_region_fresh; flowedit and ultraedit are the tier-1 rows of the same table. flowedit_t1 (FlowEdit from t = 1 at its guidance 1.5 / 5.5) is no longer a row of Tab. 2: it is in Supp. tab:fe-curve and the text of Supp. supp:design. Extras from the text and the supplement: ultraedit_g15_55 (D1), ultraedit_region_fresh_g15_55 (D2), ultraedit_no_return_no_kv (G1), ultraedit_no_return_no_kv_region_fresh (G2), flowedit_t1_g35_T16 (F2). Claims: a start at pure noise lets content move; one guidance and a region keep more of the image; noise carried on the source's forward path in the region edits more than fresh noise.

Tier 3: Fig. 3, start time and edit success (fig:exitcurve, SPRITE only; --tier 3). UltraEdit with the region's start moved later, on the forward path (ultraedit_region_start097/093/090/085) and with fresh noise (ultraedit_region_fresh_start097/...); the start-1 points are ultraedit and ultraedit_region_fresh. FlowEdit appears in Fig. 3 only as released (flowedit, n_max 24, the triangle). Its own start-time curve, n_max 28 / 26 / 25 / 22 / 20 (flowedit_t1, flowedit_n26, ...) at its guidance 1.5 / 5.5 and flowedit_t1_g35 at one guidance, is Supp. tab:fe-curve (cited in Sec. 4). Claims: edit success falls as the start moves later, and the forward path edits more than fresh noise at every start; FlowEdit's update falls at the same place.

Tier 4 is synps_inv (SynPS + inversion), kept from the first version of this repo. It is no longer in the paper; skip it unless you have already run it.

Calls per provider (each provider runs every task once):

tier	image calls	video calls	total
1	9,528	2,705	12,233
2	7,362	—	7,362
3	5,892	—	5,892
4 (optional)	491	—	491
all	23,273	2,705	25,978

Tasks whose output is byte-identical to an earlier method's output of the same edit are not sent; that answer is reused (field same_output_as). These are 225 tasks: 222 ultraedit_reanchor videos (FiVE 45, NRVBench 177) and 3 SPRITE tier-2 images (one edit each of ultraedit_whole_forward and ultraedit_outside_forward equal to ultraedit, and one of ultraedit_region_fresh equal to ultraedit_whole_fresh). Of the 12,233 tier-1 calls, 3,158 are new in this version (flowedit_t1_g35 and samflow on SPRITE and PIE-Bench, wanedit_t1_g5 on the three video benchmarks); the other 9,075 are the tasks of the first version, with the same task ids, so their answers are kept. judge.py --count prints these numbers for any selection.

Contents

path	content
images/sprite/	SPRITE: 491 edits on 207 source images (512 x 512). sources/, outputs/<method>/<edit_id>.png, edits.jsonl
images/pie/	PIE-Bench: 700 edits on 700 source images (512 x 512), same layout
videos/five/	FiVE-Bench: 419 edits on 112 source clips (100 videos)
videos/nrv/	NRVBench: 180 edits
videos/ivebench/	IVEBench motion subset: 59 edits, 3 seeds per method (<edit_id>_s0/1/2.mp4)
tasks/images.jsonl	23,276 image tasks (SPRITE 491 edits x 36 methods, PIE-Bench 700 edits x 8 methods)
tasks/videos.jsonl	2,927 video tasks (FiVE 419 x 4, NRVBench 180 x 4, IVEBench 59 x 3 methods x 3 seeds)
scripts/	judge.py (calls the API), summarize.py (success rates by paper table), requirements.txt
reference/methods.csv	every method: tier, paper row, settings, where the outputs come from, the paper's number
reference/qwen_sprite_edit_pass.csv	the paper's Qwen2.5-VL edit-pass decisions on SPRITE (every SPRITE method), for judge agreement
reference/qwen_pie_vqa.csv	the paper's Qwen2.5-VL VQA answer on PIE-Bench (every PIE-Bench method; pass = P(yes) >= 0.5)

Each line of a tasks file is one API call. It holds the source and output paths, the instruction, the success question(s) (check), tier and paper_row. For PIE-Bench, which has no such questions, check is its VQA question (Does this figure show "<target prompt>"?). Videos are 832 x 480 with 41 frames (a few FiVE clips have 25-37), mostly at 16 fps. The source videos are H.264 encodings of the exact frames the methods received.

Methods

Images (FLUX.1-dev; SPRITE seed 0, PIE-Bench seed 42). The Qwen numbers are the paper's (Edit = Qwen2.5-VL 7B + 32B edit pass on SPRITE; VQA on PIE-Bench); the reference CSVs reproduce them exactly.

tier	method	in the paper	SPRITE Edit	PIE VQA
1	ultraedit	UltraEdit (ours)	0.780	70.0
1	ultraedit_box	UltraEdit, CLIPSeg box, no return	0.792	65.1
1	flowedit_t1_g35	FlowEdit started at t = 1 with one guidance 3.5 / 3.5 (measurement; Supp. tab:fe-curve, tab:pie)	0.784	69.8
1	flowedit	FlowEdit	0.501	59.4
1	rf_solver	RF-Solver-Edit	0.487	64.1
1	fireflow	FireFlow	0.438	63.4
1	syncsde	Sync-SDE	0.405	52.7
1	samflow	SAM-Flow (official release, FLUX defaults)	0.389	44.5
2	flowedit_t1	Supp. tab:fe-curve: start at pure noise (FlowEdit n_max = T = 28, 1.5 / 5.5)	0.764
2	ultraedit_whole_forward	Tab. 2: UltraEdit, whole image from 1 (forward path)	0.804
2	ultraedit_whole_fresh	Tab. 2: whole image from 1 with fresh noise	0.603
2	ultraedit_region_fresh	Tab. 2: fresh noise in the region	0.637
2	ultraedit_region_annealed	Tab. 2: annealed noise in the region	0.690
2	ultraedit_region_fixed	Tab. 2: fixed noise in the region	0.772
2	ultraedit_outside_forward	Tab. 2: forward path outside the region	0.766
2	ultraedit_no_return	Tab. 2: no return at t_r	0.768
2	ultraedit_no_kv	Tab. 2: no shared keys and values	0.782
2	ultraedit_no_kv_region_fresh	Tab. 2: no shared K/V and fresh noise in the region	0.699
2	ultraedit_g15_55	D1: UltraEdit with FlowEdit's guidance 1.5 / 5.5	0.731
2	ultraedit_region_fresh_g15_55	D2: D1 + fresh noise in the region	0.664
2	ultraedit_no_return_no_kv	G1: no return, no shared K/V	0.790
2	ultraedit_no_return_no_kv_region_fresh	G2: G1 + fresh noise in the region	0.725
2	flowedit_t1_g35_T16	F2: flowedit_t1_g35 with 16 steps	0.758
3	ultraedit_region_start097 / 093 / 090 / 085	Fig. 3: region start t = 0.966 / 0.929 / 0.891 / 0.849, forward path	0.737 / 0.613 / 0.477 / 0.336
3	ultraedit_region_fresh_start097 / 093 / 090 / 085	Fig. 3: the same starts, fresh noise	0.572 / 0.422 / 0.273 / 0.161
3	flowedit_n26 / n25 / n22 / n20	Supp. tab:fe-curve: FlowEdit with n_max 26 / 25 / 22 / 20 (t = 0.961 / 0.940 / 0.873 / 0.824)	0.705 / 0.595 / 0.275 / 0.157
4	synps_inv	not in the current paper (SynPS + inversion)	0.735

Videos (Wan2.1-T2V-1.3B), all tier 1:

method	in the paper
ultraedit	UltraEdit (ours); seed 0, on IVEBench seeds 0-2
ultraedit_reanchor	UltraEdit + re-anchoring (FiVE and NRVBench only)
wan_edit	Wan-Edit. FiVE: the outputs released by FiVE-Bench. NRVBench and IVEBench: our run of the released code (seed 0; IVEBench seeds 0-2)
wanedit_t1_g5	Wan-Edit's released code started at pure noise (--skip_timesteps 0) with one guidance 5 / 5 (seed 0; IVEBench seeds 0-2). The paper's numbers for this row come from the benchmarks' own evaluations (FiVE-Acc 54.12 is FiVE's official evaluate.py, NRVE-Acc 32.34 NRVBench's Qwen judges, IVEBench AS 0.288 Qwen3-VL-32B), so there is no Qwen reference file for videos. In the paper: Supp. tab:supp-wan-t1
wanedit_t1_g510	Wan-Edit's released code from t = 1 with its released guidance 5 / 10. Not included: this row is not in the paper; its tasks would be added in an update of this repo (same names, same layout) if it is

reference/methods.csv lists each method's exact settings and source.

Steps

1. Download and install (Python 3.9+). The whole repo is about 17 GB. If you downloaded an earlier version into the same folder, the same command fetches only the new and changed files.

   pip install -U huggingface_hub
   hf auth login                     # an account with access to this repo
   hf download Oshikaka/ultraedit-api-judge --repo-type dataset --local-dir ultraedit-api-judge
   cd ultraedit-api-judge
   pip install -r scripts/requirements.txt
   

2. Set the keys and choose one model per provider. Use the newest flagship model with image input, and use the same model id and settings for every task and every tier. Write the ids down.

   export OPENAI_API_KEY=...  GEMINI_API_KEY=...
   GPT=<OpenAI model id>; GEM=<Gemini model id>
   python scripts/judge.py --count --tasks tasks/images.jsonl tasks/videos.jsonl    # calls per tier, no API call
   

3. Smoke test. --tier 1 --limit 8 is one SPRITE edit with all 8 tier-1 image methods, and --tier 1 --limit 4 on the videos is one FiVE edit with all 4 video methods. Open results/<provider>__<model>/*.jsonl and check the answers. --dry_run prints the exact prompt without calling the API.

   python scripts/judge.py --provider openai --model $GPT --tasks tasks/images.jsonl --tier 1 --limit 8
   python scripts/judge.py --provider gemini --model $GEM --tasks tasks/videos.jsonl --tier 1 --limit 4
   

4. Full run, tier by tier. Answers are appended as they come in. Re-running the same command skips finished tasks and retries failed ones, so repeat it until it prints failed 0. If you hit rate limits, lower --workers (default 8). If the same few tasks still fail after 2-3 re-runs (for example a refusal), stop there: summarize.py counts them as missing and also reports each row on the edits answered for all rows of its table. Answers from an earlier run of this repo are kept (task ids are unchanged), so only the new tasks are sent.

   for P in "openai $GPT" "gemini $GEM"; do set -- $P
     python scripts/judge.py --provider $1 --model $2 --tasks tasks/images.jsonl tasks/videos.jsonl --tier 1
   done
   # then --tier 2, then --tier 3 (SPRITE images only); --methods a,b selects single methods
   

5. Summarize. This writes results/summary.md, grouped by paper table, with the matched-success comparison.

   python scripts/summarize.py
   

6. Send back the whole results/ folder (the answer files and summary.md) and the two model ids, for example as a zip. Every call is one line in the answer files, written as soon as it returns: the three verdicts, the reason, the model's full answer, token usage, the model version the API reports, the response id, a hash of the prompt and all settings. Nothing else is needed to recompute the tables, so please do not delete or edit these files.

Protocol

• Input.
  • Image task: the source and the edited image (512 x 512, sent as PNG).
  • Video task: 8 frames sampled uniformly from the source video and 8 frames at the same time points from the edited video (resized to 512 px wide, JPEG). Both providers get the same frames.
  • Text: the edit instruction and the benchmark's success question(s). The prompts are IMAGE_PROMPT and VIDEO_PROMPT in judge.py. Please do not change them.
• Output. One call per task, at the provider's default temperature. The judge returns JSON {edit_success, preserved, artifacts, reason}.
• Metrics (summarize.py):
  • Success is the fraction of edits with edit_success. On IVEBench each edit is first averaged over its 3 seeds.
  • Strict is edit_success and preserved and not artifacts; preserved and artifacts are also reported on their own.
  • All come with a 95 % bootstrap interval over edits, and success with the paired difference to ultraedit.
  • Byte-identical outputs are judged once (see above). Most are ultraedit_reanchor videos (FiVE 45, NRVBench 177), whose re-anchoring leaves motion edits unchanged by design.
  • SPRITE Tab. 1 is also reported on the 420 edits not written for the paper; SPRITE and PIE-Bench also as agreement with the paper's Qwen judge.
• Size. 25,487 calls per provider for tiers 1-3 (25,978 with tier 4). An image call sends 2 images, a video call 16 frames. Optional request fields (for example a reasoning-effort setting) can be passed with --extra '<json>'. Each answer line records them; use the same fields for all tasks.