Hugging Face's logo

Datasets:
Rapidata
/
multi-step-image-edit

Modalities:
Image
Text
Formats:
parquet
Size:
10K - 100K
Tags:
image-editing
human-preference
benchmark
identity
Libraries:
Datasets
Dask
Polars
License:
Dataset card Data Studio Files
xet
 Community
7
multi-step-image-edit / metadata /models.json
s-aniruddha's picture
s-aniruddha
Card, photo credits, model settings, standings and settings evidence for the 19 ranked models
23ec548 verified
Raw History Blame
18.4 kB
{
"_rules": {
"principle": "every setting is the provider's own API default unless it is one of the interventions below (decided with Ani 2026-10-30)",
"shape": "aspect ratio 3:2 / 2:3 / 1:1 if the model has it; else its keep-aspect option; else exact size of that step's input",
"watermark": "off wherever there is a switch",
"effort": "quality / thinking / reasoning / precision at their maximum; resolution tiers stay at default; open-weights dials (Qwen-Image-Edit-2511 steps, acceleration, guidance) at fal's default (top effort tested, kept the face less well; Ani 2026-11-04)",
"moderation": "least restrictive setting we can send, incl. fal's safety checker off; moderation we cannot reach stays at the provider default; a refusal counts as a failed try",
"format": "the format the lab's own API returns when no format is requested, not fal's default (Ani 2026-11-05): JPEG for Seedream 5.0 Pro / Lite (BytePlus default jpeg) and Grok Imagine Image 2.0 (xAI returns JPEG), WebP for Muse Image (Meta default webp), PNG for the rest (OpenAI default png; Google returns PNG unless JPEG is requested; BFL, Ideogram, Microsoft, Kling and Alibaba have no format setting and return PNG; Qwen-Image-Edit-2511 has open weights, fal returns PNG). fal's output_format is set to that format. Checked on the files 2026-11-05: content credentials (C2PA) signed by OpenAI, Google, BytePlus, Ideogram, Microsoft and xAI, Kling's and Alibaba's AIGC labels and Meta's XMP are intact, and no PNG shows a JPEG block pattern except fal's PNG copy of Grok's JPEG",
"prompt_expansion": "provider's documented default where there is one; off where the provider has a setting without a documented default or has no expansion; fal-side expansion switches always off",
"seed": "int(sha256('model|photo|step|try')) mod 2^31 where the model takes a seed",
"request": "one image per call; the instruction text is the only prompt; every setting sent explicitly",
"fal_only_plumbing": "settings with no provider equivalent pinned to: one image per call, whole-image edit, opaque background",
"chain": "step 1 gets the reference PNG; each later step gets the previous output unchanged (its own URL), whatever format and size; raw files kept",
"check": "gpt-6.1-sol (medium) sees the raw before/after; partial accepted; 3 tries; chain stops after 3 failed tries; timeouts and server errors do not use a try",
"crop": "step-5 output resized to the reference size (Lanczos), same face box (1.7x largest face, centre raised 5%), scaled to 768 px, saved as PNG",
"exclusion": "aspect ratio changed by more than 1%; models below 80% step-5 completion are not voted (Ani 2026-11-03; Grok Imagine Image (quality) allowed just under)",
"record": "fal_schemas/ and provider_docs/ saved 2026-10-30 (Wayback links in provider_docs/fetch_log.tsv and wayback_retry.tsv); each try's final request and response logged by the runner (retries after timeouts and server errors are not)",
"calls": "fal sync endpoint (fal.run) since 2026-11-03 after fal's queue stalled; Alibaba DashScope direct for Qwen-Image-3.0-Pro and -2.1-Pro; Replicate for P-Image-Edit"
},
"approved": "2026-10-30",
"updated": "2026-11-05",
"models": [
{
"name": "gpt-image-2.5-sunburst",
"label": "GPT Image 2.5 Sunburst (Max Quality)",
"provider": "fal",
"endpoint": "openai/gpt-image-2.5/sunburst/edit",
"image_field": "image_urls",
"shape": "fixed",
"send": {
"image_size": "auto",
"quality": "max",
"background": "auto",
"output_format": "png",
"num_images": 1,
"enable_safety_checker": false
},
"seed": false,
"source": "OpenAI images/edits: output_format default png, quality default auto (max exists), size default auto, background auto, moderation auto (low not exposed on fal)",
"status": "on_leaderboard",
"participant_id": "par_1ZqchVbbZcVgKP",
"list_price_per_image_usd": 0.18,
"price_basis": "re-measured 2026-11-05 like GPT Image 2.6: 5,488 output tokens (3:2, 2:3), 8,232 (1:1) at quality max; was $0.154 from the 2026-10-27 measurement (4,667 / 7,024 tokens)",
"reached_step5": "196/199"
},
{
"name": "gpt-image-2.5-flare",
"label": "GPT Image 2.5 Flare (Max Quality)",
"provider": "fal",
"endpoint": "openai/gpt-image-2.5/flare/edit",
"image_field": "image_urls",
"shape": "fixed",
"send": {
"image_size": "auto",
"quality": "max",
"background": "auto",
"output_format": "png",
"num_images": 1,
"enable_safety_checker": false
},
"seed": false,
"source": "as Sunburst",
"status": "on_leaderboard",
"participant_id": "par_1ZqciFXwVbmuqF",
"list_price_per_image_usd": 0.18,
"price_basis": "re-measured 2026-11-05 like GPT Image 2.6: 5,488 output tokens (3:2, 2:3), 8,232 (1:1) at quality max; was $0.154 from the 2026-10-27 measurement (4,667 / 7,024 tokens)",
"reached_step5": "191/199"
},
{
"name": "nano-banana-2.1",
"label": "Nano Banana 2.1 (High Thinking)",
"provider": "fal",
"endpoint": "google/nano-banana-2.1/edit",
"image_field": "image_urls",
"shape": "aspect_ratio",
"send": {
"thinking_level": "high",
"resolution": "1K",
"output_format": "png",
"safety_tolerance": "6",
"enable_web_search": false,
"limit_generations": true,
"num_images": 1
},
"seed": true,
"source": "Gemini image generation: 1K default, thinking always on (level model-dependent), SynthID always on, png/jpeg both supported with no stated default (-> png rule), seed optional",
"status": "on_leaderboard",
"participant_id": "par_1Zqcin3reEdVrL",
"list_price_per_image_usd": 0.043,
"price_basis": "Google: 1,120 out tok x $30/M + 1,120 in x $1.50/M + ~1,000 thinking x $7.50/M",
"reached_step5": "192/199"
},
{
"name": "nano-banana-pro",
"label": "Nano Banana Pro",
"provider": "fal",
"endpoint": "fal-ai/nano-banana-pro/edit",
"image_field": "image_urls",
"shape": "aspect_ratio",
"send": {
"resolution": "1K",
"output_format": "png",
"safety_tolerance": "6",
"enable_web_search": false,
"limit_generations": true,
"num_images": 1
},
"seed": true,
"source": "as Nano Banana 2.1; thinking always on with a single level, so not sent",
"status": "on_leaderboard",
"participant_id": "par_1ZqcjI3Jy4BUCH",
"list_price_per_image_usd": 0.148,
"price_basis": "Google: 1,120 out tok x $120/M + 560 in x $2/M + ~1,000 thinking x $12/M",
"reached_step5": "185/199"
},
{
"name": "seedream-5-pro",
"label": "Seedream 5.0 Pro",
"provider": "fal",
"endpoint": "bytedance/seedream/v5/pro/edit",
"image_field": "image_urls",
"shape": "fixed",
"send": {
"image_size": "auto_2K",
"watermark": false,
"output_format": "jpeg",
"enhance_prompt_mode": "standard",
"background": "opaque",
"num_images": 1,
"enable_safety_checker": false
},
"seed": false,
"source": "BytePlus ModelArk: watermark default true (-> off, intervention), output_format default jpeg, optimize_prompt mode default standard, size 2K",
"status": "on_leaderboard",
"participant_id": "par_1ZqcjmwxM2fP66",
"list_price_per_image_usd": 0.09,
"price_basis": "BytePlus ModelArk >2.61 MP tier $0.09/image (v2 outputs 4.15 MP at the 2K default)",
"reached_step5": "192/199"
},
{
"name": "seedream-5-lite",
"label": "Seedream 5.0 Lite",
"provider": "fal",
"endpoint": "bytedance/seedream/v5/lite/edit",
"image_field": "image_urls",
"shape": "fixed",
"send": {
"image_size": "auto_2K",
"watermark": false,
"output_format": "jpeg",
"num_images": 1,
"max_images": 1,
"enable_safety_checker": false
},
"seed": false,
"source": "BytePlus ModelArk: watermark default true (-> off), output_format jpeg, prompt optimisation standard (fast unsupported), default size 2048x2048 square (-> keep-aspect auto_2K, shape intervention), min 2560x1440",
"status": "on_leaderboard",
"participant_id": "par_1ZqckGxhfElEwY",
"list_price_per_image_usd": 0.035,
"price_basis": "BytePlus ModelArk $0.035/image, input free",
"reached_step5": "188/199"
},
{
"name": "muse",
"label": "Muse Image",
"provider": "fal",
"endpoint": "meta/muse-image/edit",
"image_field": "image_urls",
"shape": "aspect_ratio",
"send": {
"output_format": "webp",
"num_images": 1
},
"seed": false,
"source": "Meta images API: output_format default webp, reasoning_strength default high (max; not exposed on fal), moderation auto/low (not exposed on fal), no seed",
"status": "on_leaderboard",
"participant_id": "par_1ZqckkPfwoUxSz",
"list_price_per_image_usd": 0.01,
"price_basis": "Meta Model API flat $0.01/image",
"reached_step5": "190/199"
},
{
"name": "flux-3",
"label": "FLUX 3 Image",
"provider": "fal",
"endpoint": "blackforestlabs/flux-3/edit-image",
"image_field": "image_urls",
"shape": "aspect_ratio",
"send": {
"resolution": "1k",
"output_format": "png",
"safety_tolerance": 4,
"enable_prompt_expansion": false
},
"seed": false,
"source": "BFL openapi Flux3ImageInputs: no format, seed or expansion field; resolution 1k; safety_tolerance default 2 on 0-7 (fal max 4); grounding true; docs: FLUX 3 expands every instruction itself",
"status": "on_leaderboard",
"participant_id": "par_1ZqclEHD4AV5Lx",
"list_price_per_image_usd": 0.048,
"price_basis": "BFL 1k tier 4.8c, reference images free",
"reached_step5": "193/199"
},
{
"name": "ideogram-4.5",
"label": "Ideogram 4.5 (High Quality, High Precision)",
"provider": "fal",
"endpoint": "ideogram/v4.5/edit",
"image_field": "image_url",
"shape": "fixed",
"send": {
"image_size": "auto",
"quality": "high",
"edit_precision": "high",
"context_window": "none",
"num_images": 1,
"enable_safety_checker": false
},
"seed": true,
"source": "Ideogram openapi precise-edit/ideogram-4-5: quality default medium (max high), seed supported, no magic_prompt on precise edit; no format setting",
"status": "on_leaderboard",
"participant_id": "par_1ZqcliNedxU86d",
"list_price_per_image_usd": 0.22,
"price_basis": "Ideogram API quality high $0.22/image",
"reached_step5": "195/199"
},
{
"name": "mai-image-2.5-pro",
"label": "MAI-Image-2.5-Pro",
"provider": "fal",
"endpoint": "microsoft/mai-image-2.5-pro/edit",
"image_field": "image_url",
"shape": "aspect_ratio",
"send": {
"output_format": "png",
"num_images": 1
},
"seed": false,
"source": "Microsoft Foundry MAI: output always PNG, edit inputs JPEG/PNG, no size field on edits, 2.5 max 1,048,576 px",
"status": "on_leaderboard",
"participant_id": "par_1ZqcmCtOtXcinz",
"list_price_per_image_usd": 0.117,
"price_basis": "Microsoft Foundry $106/M out, $8/M image in; ~1,024 tok/MP inferred",
"reached_step5": "172/199"
},
{
"name": "grok-imagine-2.0",
"label": "Grok Imagine Image 2.0 (Medium Quality)",
"provider": "fal",
"endpoint": "xai/grok-imagine-image/v2.0/edit",
"image_field": "image_urls",
"shape": "aspect_ratio",
"send": {
"quality": "medium",
"resolution": "1k",
"output_format": "jpeg",
"num_images": 1,
"enable_safety_checker": false
},
"seed": false,
"source": "xAI images: quality low/medium/auto default auto (medium = max), resolution 1k, mime jpeg unless quality produces png, aspect default auto",
"status": "on_leaderboard",
"participant_id": "par_1ZqcofFmAZr4Yi",
"list_price_per_image_usd": 0.04,
"price_basis": "xAI pricing page, flat $0.04/image",
"reached_step5": "175/199"
},
{
"name": "kling-o1",
"label": "Kling Image O1",
"provider": "fal",
"endpoint": "fal-ai/kling-image/o1",
"image_field": "image_urls",
"shape": "aspect_ratio",
"send": {
"resolution": "1K",
"output_format": "png",
"num_images": 1,
"enable_safety_checker": false
},
"seed": false,
"source": "Kling omni image API: model_name default kling-image-o1 (this model), resolution 1k, aspect auto, watermark_info optional separate copy, no format field (-> png rule). Chosen over Kling Image 3.0 Omni because fal's o3 endpoint does not report which model runs (Ani, 2026-10-30)",
"status": "on_leaderboard",
"participant_id": "par_1Zqcme5IPaYe28",
"list_price_per_image_usd": 0.028,
"price_basis": "Kling API 8 units x $0.0035",
"reached_step5": "167/199"
},
{
"name": "wan-2.7-pro",
"label": "Wan 2.7 Pro",
"provider": "fal",
"endpoint": "fal-ai/wan/v2.7/pro/edit",
"image_field": "image_urls",
"shape": "exact",
"send": {
"output_format": "png",
"enable_prompt_expansion": false,
"num_images": 1,
"enable_safety_checker": false
},
"seed": true,
"source": "Alibaba Wan image API: PNG, watermark default false, edits 2K at input aspect, seed optional, no prompt_extend listed for wan2.7-image-pro",
"status": "on_leaderboard",
"participant_id": "par_1ZqcnCQjqEzo3J",
"list_price_per_image_usd": 0.075,
"price_basis": "Alibaba Model Studio wan2.7-image-pro $0.075, output only",
"reached_step5": "178/199"
},
{
"name": "qwen-image-3.0-pro",
"label": "Qwen-Image-3.0-Pro (Thinking)",
"provider": "dashscope",
"endpoint": "qwen-image-3.0-pro",
"image_field": null,
"shape": "exact",
"send": {
"n": 1,
"watermark": false,
"prompt_extend": true,
"enable_thinking": true
},
"seed": true,
"source": "Alibaba Model Studio qwen-image API: prompt_extend default true, enable_thinking default true, watermark default false, PNG output, size w*h (default model-picked), seed optional",
"status": "on_leaderboard",
"participant_id": "par_1Zqcnfvh0ICqlR",
"list_price_per_image_usd": 0.043,
"price_basis": "Alibaba Model Studio (Singapore) 1K $0.04 output + $0.003 input image",
"reached_step5": "191/199"
},
{
"name": "qwen-image-2.1-pro",
"label": "Qwen-Image-2.1-Pro (Thinking)",
"provider": "dashscope",
"endpoint": "qwen-image-2.1-pro",
"image_field": null,
"shape": "exact",
"send": {
"n": 1,
"watermark": false,
"prompt_extend": true,
"enable_thinking": true
},
"seed": true,
"source": "Alibaba Model Studio qwen-image API: prompt_extend default true, enable_thinking default true (requires prompt_extend), watermark default false, PNG output, size w*h, seed optional",
"status": "on_leaderboard",
"participant_id": "par_1ZqcoB8cEAbf76",
"list_price_per_image_usd": 0.04,
"price_basis": "Alibaba Model Studio (Singapore) $0.04/image",
"reached_step5": "191/199"
},
{
"name": "qwen-image-edit-2511",
"label": "Qwen-Image-Edit-2511",
"provider": "fal",
"endpoint": "fal-ai/qwen-image-edit-2511",
"image_field": "image_urls",
"shape": "exact",
"send": {
"num_images": 1,
"output_format": "png",
"enable_safety_checker": false
},
"seed": true,
"source": "open weights, no lab API: fal schema; steps 28, acceleration regular, guidance 4.5 left at default (v2 rule: open-weights dials at default); png default; checker off",
"status": "on_leaderboard",
"participant_id": "par_1ZrTY2UFNLac98",
"list_price_per_image_usd": 0.03,
"price_basis": "Alibaba Model Studio qwen-image-edit-plus $0.03/image (open weights)",
"reached_step5": "160/199"
},
{
"name": "nano-banana-2",
"label": "Nano Banana 2 (High Thinking)",
"provider": "fal",
"endpoint": "fal-ai/nano-banana-2/edit",
"image_field": "image_urls",
"shape": "aspect_ratio",
"send": {
"thinking_level": "high",
"resolution": "1K",
"output_format": "png",
"safety_tolerance": "6",
"enable_web_search": false,
"limit_generations": true,
"num_images": 1
},
"seed": true,
"source": "as Nano Banana 2.1: Gemini 1K default, thinking high = max, png (no Google default), safety_tolerance 6 = least restrictive, web search off",
"status": "on_leaderboard",
"participant_id": "par_1ZrTYu0JqTTp2g",
"list_price_per_image_usd": 0.071,
"price_basis": "Google: 1,120 out tok x $60/M + input + ~1,000 thinking x $3/M",
"reached_step5": "183/199"
},
{
"name": "gpt-image-2.6-sunburst",
"label": "GPT Image 2.6 Sunburst (Max Quality, High Reasoning)",
"provider": "fal",
"endpoint": "openai/gpt-image-2.6/sunburst/edit",
"image_field": "image_urls",
"shape": "fixed",
"send": {
"image_size": "auto",
"quality": "max",
"reasoning_effort": "high",
"background": "auto",
"output_format": "png",
"num_images": 1,
"enable_safety_checker": false
},
"seed": false,
"source": "fal schema 2026-11-05: quality default high (max exists), reasoning_effort low/medium/high default high (= max), output_format default png, image_size auto keeps the input shape, background auto (OpenAI default); no seed; checker off",
"status": "on_leaderboard",
"list_price_per_image_usd": 0.18,
"price_basis": "OpenAI token rates ($30/M image out, $8/M image in, $5/M text in) × measured usage at our settings: 5,488 output tokens (3:2, 2:3), 8,232 (1:1); weighted 182/17",
"reached_step5": "196/199",
"participant_id": "par_1a2UkRyJ71D4J8"
},
{
"name": "gpt-image-2.6-flare",
"label": "GPT Image 2.6 Flare (Max Quality, High Reasoning)",
"provider": "fal",
"endpoint": "openai/gpt-image-2.6/flare/edit",
"image_field": "image_urls",
"shape": "fixed",
"send": {
"image_size": "auto",
"quality": "max",
"reasoning_effort": "high",
"background": "auto",
"output_format": "png",
"num_images": 1,
"enable_safety_checker": false
},
"seed": false,
"source": "fal schema 2026-11-05: quality default high (max exists), reasoning_effort low/medium/high default high (= max), output_format default png, image_size auto keeps the input shape, background auto (OpenAI default); no seed; checker off",
"status": "on_leaderboard",
"list_price_per_image_usd": 0.18,
"price_basis": "OpenAI token rates ($30/M image out, $8/M image in, $5/M text in) × measured usage at our settings: 5,488 output tokens (3:2, 2:3), 8,232 (1:1); weighted 182/17",
"reached_step5": "194/199",
"participant_id": "par_1a2UlVdJSE8VFt"
}
]
}