DHM · fal.ai i2i bake-off · 2026-08-05

Image-edit models on the v7.2 held-out set

Each v7.2 showcase preview (1600 px) sent to seven fal.ai image-edit endpoints, same instruction, seed 42 where supported. Arm A = all 28 scenes, neutral enhancement prompt. Arm B = the 10 worst-locality scenes (per the 2026-08-05 white-gap probe), prompt targeting the three live residuals: wall evenness, cast neutralization, highlight protection. struct % = pixels changed after tone-normalization (content drift — lower is safer). edge Δ% = edge-map flips. wall ΔL = wall luminance vs the human-edited beat (0 = editor-level). Input card is outlined. Click any card for full size.

Arm A: Enhance this real estate photograph: brighter, neutral white balance, professional magazine look. Do not add, remove, move, or restyle any objects or furniture. Do not alter, add, or remove any text.
Arm B: Retouch this real estate interior photograph like a professional photo editor. Even out the lighting on walls and ceilings: each flat painted surface should be uniformly lit, with no hot spots, bright patches, or shadowed corners. Neutralize any color cast on white walls, ceilings, and trim so they read as clean neutral white. Keep the overall exposure unchanged - do not brighten or darken the image globally. Protect the highlights: window views must stay detailed, never washed out or blown to white. Photographic tone correction only: do not add, remove, move, or restyle any objects, furniture, or fixtures; do not change materials or textures; do not alter, add, or remove any text.

→ trained-LoRA results moved to the LoRA lab page

Arm A summary — neutral prompt, 28 scenes

modelnstruct % ↓edge Δ% ↓|wall ΔL| vs beat ↓wall chroma (beat≈2.9)
Qwen Image Edit 25112714.712.56.83.5
FLUX.2 edit278.012.412.44.7
Kontext Max2712.811.52.94.5
Kontext dev277.28.14.35.1
Nano Banana 2271.66.05.42.9
Nano Banana Pro274.17.96.03.1
Seedream 5 lite2711.013.014.62.2

Arm B summary — targeted prompt, 10 locality-worst scenes

modelnstruct % ↓edge Δ% ↓|wall ΔL| vs beat ↓wall chroma (beat≈2.9)
Qwen Image Edit 25111016.612.53.23.9
FLUX.2 edit1010.410.66.64.9
Kontext Max1011.310.02.34.1
Kontext dev102.95.62.44.5
Nano Banana 2102.36.03.13.5
Nano Banana Pro105.38.23.23.3
Seedream 5 lite109.812.43.43.5

Arm C summary — trained FLUX.2 edit LoRA, same 10 scenes

modelnstruct % ↓edge Δ% ↓|wall ΔL| vs beat ↓wall chroma (beat≈2.9)
FLUX.2 + DHM LoRA (1k steps)108.011.03.04.6

Arm D summary — trained LoRA, holdout property (11 Scotch Pine Cir, non-guide renders, unseen in training)

modelnstruct % ↓edge Δ% ↓|wall ΔL| vs beat ↓wall chroma (beat≈2.9)
FLUX.2 + DHM LoRA (1k steps)107.110.93.511.6

15-Searle-Rd · scene 002

Arm A — neutral prompt

15-Searle-Rd · scene 004

Arm A — neutral prompt

Arm B — targeted prompt

Arm C — trained FLUX.2 edit LoRA (61 pairs, 1k steps, caption prompt)

16-Denton-Rd · scene 018

Arm A — neutral prompt

16-Denton-Rd · scene 023

Arm A — neutral prompt

Arm B — targeted prompt

Arm C — trained FLUX.2 edit LoRA (61 pairs, 1k steps, caption prompt)

19-New-Haven-St · scene 007

Arm A — neutral prompt

19-New-Haven-St · scene 012

Arm A — neutral prompt

200-Royal-Street · scene 011

Arm A — neutral prompt

Arm B — targeted prompt

Arm C — trained FLUX.2 edit LoRA (61 pairs, 1k steps, caption prompt)

200-Royal-Street · scene 013

Arm A — neutral prompt

200-Royal-Street · scene 022

Arm A — neutral prompt

26-Driftwood-Cir · scene 018

Arm A — neutral prompt

26-Driftwood-Cir · scene 031

Arm A — neutral prompt

26-Driftwood-Cir · scene 032

Arm A — neutral prompt

33-Charter-St · scene 025

Arm A — neutral prompt

34R-Clyde-St · scene 013

Arm A — neutral prompt

34R-Clyde-St · scene 018

Arm A — neutral prompt

40-Chatanika-Ave · scene 006

Arm A — neutral prompt

Arm B — targeted prompt

Arm C — trained FLUX.2 edit LoRA (61 pairs, 1k steps, caption prompt)

40-Chatanika-Ave · scene 025

Arm A — neutral prompt

Arm B — targeted prompt

Arm C — trained FLUX.2 edit LoRA (61 pairs, 1k steps, caption prompt)

470-Quinobequin-Rd · scene 021

Arm A — neutral prompt

470-Quinobequin-Rd · scene 042

Arm A — neutral prompt

Arm B — targeted prompt

Arm C — trained FLUX.2 edit LoRA (61 pairs, 1k steps, caption prompt)

574-Foundry-St · scene 009

Arm A — neutral prompt

574-Foundry-St · scene 022

Arm A — neutral prompt

67-Dudley-St · scene 001

Arm A — neutral prompt

67-Dudley-St · scene 016

Arm A — neutral prompt

Arm B — targeted prompt

Arm C — trained FLUX.2 edit LoRA (61 pairs, 1k steps, caption prompt)

75-Grove-St-Unit-227 · scene 040

Arm A — neutral prompt

903-E-Broadway-unit-2 · scene 009

Arm A — neutral prompt

Arm B — targeted prompt

Arm C — trained FLUX.2 edit LoRA (61 pairs, 1k steps, caption prompt)

903-E-Broadway-unit-2 · scene 015

Arm A — neutral prompt

Arm B — targeted prompt

Arm C — trained FLUX.2 edit LoRA (61 pairs, 1k steps, caption prompt)

903-E-Broadway-unit-2 · scene 018

Arm A — neutral prompt

Arm B — targeted prompt

Arm C — trained FLUX.2 edit LoRA (61 pairs, 1k steps, caption prompt)

11_Scotch_Pine_Cir · scene s009

Arm D — trained LoRA on HOLDOUT property (non-guide workflow renders, window-heavy)

11_Scotch_Pine_Cir · scene s010

Arm D — trained LoRA on HOLDOUT property (non-guide workflow renders, window-heavy)

11_Scotch_Pine_Cir · scene s013

Arm D — trained LoRA on HOLDOUT property (non-guide workflow renders, window-heavy)

11_Scotch_Pine_Cir · scene s018

Arm D — trained LoRA on HOLDOUT property (non-guide workflow renders, window-heavy)

11_Scotch_Pine_Cir · scene s019

Arm D — trained LoRA on HOLDOUT property (non-guide workflow renders, window-heavy)

11_Scotch_Pine_Cir · scene s022

Arm D — trained LoRA on HOLDOUT property (non-guide workflow renders, window-heavy)

11_Scotch_Pine_Cir · scene s023

Arm D — trained LoRA on HOLDOUT property (non-guide workflow renders, window-heavy)

11_Scotch_Pine_Cir · scene s028

Arm D — trained LoRA on HOLDOUT property (non-guide workflow renders, window-heavy)

11_Scotch_Pine_Cir · scene s029

Arm D — trained LoRA on HOLDOUT property (non-guide workflow renders, window-heavy)

11_Scotch_Pine_Cir · scene s032

Arm D — trained LoRA on HOLDOUT property (non-guide workflow renders, window-heavy)