EvoGenRESEARCH

EVOGEN-HARNESS / TECHNICAL REPORT

EvoGen-Harness. Same model.
Smarter harness.

Learning where and how to evolve
image-generation harnesses.

Turn visual failures into targeted, reusable improvements—without updating the image generator’s weights.

Read the paper Explore examples
Frozen generator5 editable responsibilities
A LOCAL CHANGE. A VISIBLE DIFFERENCE.03 / TOOLS
Before
Before harness evolution: a calico cat with a red bow tie.
Attribute mismatch
With EvoGen
After harness evolution: a calico cat with a navy polka-dot bow tie.
Targeted harness update
REQUEST DETAIL

“A calico cat wearing a navy polka-dot bow tie…”

PolicyToolsSkillsMiddlewareMemory
Selected paper example · explore the full prompt

Jiabin Luo1Yinan Liu2Chunlei Meng3Yufei Guo1,*

1Peking University2Beijing University of Technology3Fudan University

* Corresponding author

GenEval2+0.2633Overall Soft-TIFA GM T2I-CompBench+++0.0720Average compositional score WISE+0.0752Overall WiScore

Absolute score gains over the strongest evaluated baseline in each benchmark. Reported in Tables 1–3; these are not percentage gains.

01 / SEE THE DIFFERENCE

Same prompt.
A more capable system.

Different failures call for different changes.
Explore four targeted updates to the external harness.

POLICY / COUNTING

Four means four.

Before evolutionParent harness
Count constraint missed
With EvoGen-HarnessEvolved harness
Count constraint repaired
ORIGINAL PROMPT

A red vintage tram crossing a rainy city square at dusk, with four cyclists in reflective yellow jackets riding behind it, and three glowing street lanterns along the left side.

WHAT CHANGESVisual requirements

Strengthen an explicit counting constraint in the persistent requirements, rather than changing the generator.

Report, Appendix B.6 · Figure 6(a)

Selected qualitative examples from the report, not an exhaustive evaluation. Original image contents are preserved. Switching examples changes the display; it does not run a model. Middleware is documented in the method and the report’s quantitative analysis.

02 / PROGRESSIVE REPAIR

One correction isn’t
always enough.

After each update, TRACE re-examines the remaining failures—so evolution can move to another responsibility.

Report Figure 4

“A panda sitting on a wooden bench with two red balloons floating on the right side, in a park.”

01Initial output
Count missedPosition missed
02After Policy edit
Count repairedPosition missed
03After further evolution
Count repairedPosition repaired

The generator remains frozen across the sequence. The final output may involve additional responsibility-specific updates; these examples do not imply a single fixed edit sequence for every prompt.

03 / INSIDE THE METHOD

Learn where to evolve.
Then learn how.

Visual feedback tells us what failed.
It does not, by itself, tell us what to change.

TRACE connects repeated visual evidence to the external decisions that produced it, then explores localized updates.

THE EDITABLE HARNESSImage generator: frozen
VISUAL REQUIREMENTS

What should be satisfied?

Keeps the interpretation of requested objects, attributes, counts, and relations explicit. A localized edit can strengthen a requirement that the current harness repeatedly misses.

91.4%Attribution recallControlled analysis · Table 6
TRACE

Trajectory-Relative Attribution and Coordinated Evolution

01

Collect evidence

Aggregate constraint checks over multiple stochastic executions.

02

Localize

Keep a compact frontier of plausible responsibilities—or choose No-Patch.

03

Propose & compare

Explore localized edits under matched generation conditions.

04

Re-attribute

After a useful update, localize the residual failures again.

05

Validate & retain

Keep improvements that pass held-out and preservation checks; otherwise roll back.

THE FULL FRAMEWORKVector PDF
Figure 1 from the report. Click to enlarge.

04 / BEYOND SELECTED EXAMPLES

Measured improvements.
Explicit trade-offs.

Results across instruction following, composition, and world knowledge. Scores below are reported in the technical report.

OVERALL SOFT-TIFA GM

GenEval2

0.7089+0.2633 vs. best evaluated baseline

Fine-grained instruction following across objects, attributes, counting, position, and verbs. Overall scores use prompt-level Soft-TIFA GM.

EvoGen-HarnessOurs
0.7089
Nano Banana
0.4456
Qwen-Image
0.3483
Z-Image
0.3034
OctoT2I
0.2895
Top five reported scores · higher is betterTable 1
All evaluated methods and scores
GenEval2 · Report Table 1
MethodOverall score ↑
EvoGen-Harness0.7089
Nano Banana0.4456
Qwen-Image0.3483
Z-Image0.3034
OctoT2I0.2895
VisualPrompter0.2552
Flow-GRPO0.2198
FLUX.1-dev0.2115
Promptist0.2056
NeuroPrompts0.1976
Janus-Pro0.1878
BeautifulPrompt0.1134
Idea2Img0.0586
GenArtist0.0556
ChatGen0.0388
Download benchmark scores (CSV)

Default EvoGen-Harness generator: FLUX.1-dev. Baseline configurations follow the report; different standalone models are also evaluated. This is a summary of the manuscript, not an independently reproduced or continuously updated leaderboard.

QUALITY IS NOT THE ONLY AXIS

In the reported GenEval2 inference setting, EvoGen-Harness takes 42.6 s per prompt versus OctoT2I’s 10.0 s, with 1.4 vs. 1.3 generator calls. Evolution-time search is a separate cost. See Tables 5 and 9 for the full accounting

THREE FROZEN BACKBONES

One framework. Different generators.

FLUX.1-dev

GenEval2 overall

0.21150.7089

Qwen-Image

GenEval2 overall

0.34830.7594

Janus-Pro

GenEval2 overall

0.18780.7002

Each harness is evolved separately for its frozen backbone; this is not zero-shot transfer of one evolved harness. Scores: Table 4. Selected images: Figure 7. View original prompts and figure

First page of the EvoGen-Harness technical report

05 / READ. REPRODUCE. BUILD.

The full story,
beyond the gallery.

EvoGen-Harness: Learning Where and How to Evolve Image-Generation Harnesses

Method, experimental settings, ablations, additional examples, and the Harness Localizer prompt—in one 23-page technical report.

Read the reportView repository
Download PDF

This site is an interactive research showcase, not a hosted image-generation service. Check the repository for implementation availability.

CITE THIS WORK

BibTeX

@article{luo2026evogen,
  title={EvoGen-Harness: Learning Where and How to Evolve Image-Generation Harnesses},
  author={Luo, Jiabin and Liu, Yinan and Meng, Chunlei and Guo, Yufei},
  journal={arXiv preprint arXiv:2610.00383},
  year={2026}
}
Download .bib

A few useful distinctions.

Does this fine-tune the image generator?

No. The framework changes persistent generator-external responsibilities while keeping the generator’s weights frozen. The report also uses fixed LLMs for attribution and edit proposal.

Is this just prompt optimization?

Prompt optimization chooses a predefined adaptation target. EvoGen-Harness opens five external responsibilities to evolution; TRACE uses execution evidence to focus where edits are explored and revisits that decision after each accepted update.

Can I enter a prompt and generate an image here?

No. The controls replay and compare existing outputs from the paper. This static website has no inference backend, API key input, or live model connection. Visit the repository for implementation status.

Does a targeted update guarantee every constraint is solved?

No. The gallery contains selected examples, not guaranteed outcomes. Correct attribution does not ensure a successful edit, and stochastic variation and verifier errors remain possible. TRACE uses No-Patch and held-out validation to limit unnecessary or harmful persistent updates. The report discusses these limitations and computational costs.

FROZEN WEIGHTS. ROOM TO EVOLVE.

Better systems,
not just bigger models.

Explore the evidence. Read the method.
Build on the research.

Explore the repository

Image preview

Open full image