# Human Preference Optimization on REFL (300 random images)

Source: [Human Preference Optimization on REFL (300 random images)](<https://www.sota2.com/research/sota/human-preference-optimization-on-refl-300-random-images>)

Collection index: [llms.txt](<https://www.sota2.com/research/llms.txt>)

Data updated: 2026-03-05T08:29:45.541Z

Evaluation of generative model performance using human preference optimization metrics (ImageReward, HPSv2, CLIPScore, Aesthetic, LPIPS) applied to the REFL dataset subset containing 300 random images.

A research result applies to its stated dataset, split, metric and evaluation configuration. Compare like-for-like settings, preserve the metric's unit and direction, and distinguish paper-reported results from independent reproduction. The best result recorded here is not a claim of exhaustive, current global coverage.

| Scope | Recorded value |
| --- | --- |
| Task | Human preference optimization |
| Dataset (including any stated split) | REFL (300 random images) |
| Metric | ImageReward |
| Unit | SCALAR |
| Direction | Higher is better |
| Results loaded | 6 of 6 |
| Coverage | All results currently returned for this benchmark |

Evaluation settings may differ between results. The original papers and recorded configurations are needed to establish comparability. A shared metric name alone does not establish an identical protocol.

## Metrics

- [ImageReward](<https://www.sota2.com/research/sota/human-preference-optimization-on-refl-300-random-images.md?metric=ImageReward>): Higher is better; unit: SCALAR.

- [HPSv2](<https://www.sota2.com/research/sota/human-preference-optimization-on-refl-300-random-images.md?metric=HPSv2>): Higher is better; unit: SCALAR.

- [CLIPScore](<https://www.sota2.com/research/sota/human-preference-optimization-on-refl-300-random-images.md?metric=CLIPScore>): Higher is better; unit: SCALAR.

- [Aesthetic Score](<https://www.sota2.com/research/sota/human-preference-optimization-on-refl-300-random-images.md?metric=Aesthetic%20Score>): Higher is better; unit: SCALAR.

- [LPIPS](<https://www.sota2.com/research/sota/human-preference-optimization-on-refl-300-random-images.md?metric=LPIPS>): Lower is better; unit: SCALAR.

- [CLIP-Isrc](<https://www.sota2.com/research/sota/human-preference-optimization-on-refl-300-random-images.md?metric=CLIP-Isrc>): Higher is better; unit: SCALAR.

## Recorded results

| Method | ImageReward | Result date | Report paper |
| --- | --- | --- | --- |
| Gradient Ascent | 1.9088 | Not recorded | 68fa7bc81e209131cea8a752 |
| Trajectory Optimization Control | 1.8914 | Not recorded | 68fa7bc81e209131cea8a752 |
| Inversion+TFG | 1.7053 | Not recorded | 68fa7bc81e209131cea8a752 |
| Inversion+FreeDoM | 1.5995 | Not recorded | 68fa7bc81e209131cea8a752 |
| Inversion+DPS | 1.5988 | Not recorded | 68fa7bc81e209131cea8a752 |
| None | 0.1542 | Not recorded | 68fa7bc81e209131cea8a752 |

### Gradient Ascent

ImageReward: 1.9088. Result ID: 69a93ee76dbb7075b952a828.

[Reporting paper](<https://www.sota2.com/research/paper/68fa7bc81e209131cea8a752.md>)

[Method origin paper](<https://www.sota2.com/research/paper/68fa61c91e209131ce9f5e06.md>)

Recorded evaluation configuration: Not recorded. This record does not assert independent reproduction.

### Trajectory Optimization Control

ImageReward: 1.8914. Result ID: 69a93ee76dbb7075b952a82c.

[Reporting paper](<https://www.sota2.com/research/paper/68fa7bc81e209131cea8a752.md>)

[Method origin paper](<https://www.sota2.com/research/paper/68fa7bc81e209131cea8a752.md>)

Recorded evaluation configuration: Not recorded. This record does not assert independent reproduction.

### Inversion+TFG

ImageReward: 1.7053. Result ID: 69a93ee76dbb7075b952a82b.

[Reporting paper](<https://www.sota2.com/research/paper/68fa7bc81e209131cea8a752.md>)

[Method origin paper](<https://www.sota2.com/research/paper/68fa75bc1e209131cea699d8.md>)

[Implementation](<https://www.sota2.com/Code%20will%20be%20available%20soon.>)

Recorded evaluation configuration: Not recorded. This record does not assert independent reproduction.

### Inversion+FreeDoM

ImageReward: 1.5995. Result ID: 69a93ee76dbb7075b952a82a.

[Reporting paper](<https://www.sota2.com/research/paper/68fa7bc81e209131cea8a752.md>)

[Method origin paper](<https://www.sota2.com/research/paper/68fa6ec71e209131cea42035.md>)

Recorded evaluation configuration: Not recorded. This record does not assert independent reproduction.

### Inversion+DPS

ImageReward: 1.5988. Result ID: 69a93ee76dbb7075b952a829.

[Reporting paper](<https://www.sota2.com/research/paper/68fa7bc81e209131cea8a752.md>)

[Method origin paper](<https://www.sota2.com/research/paper/68fa7bdd1e209131cea8adbc.md>)

Recorded evaluation configuration: Not recorded. This record does not assert independent reproduction.

### None

ImageReward: 0.1542. Result ID: 69a93ee76dbb7075b952a827.

[Reporting paper](<https://www.sota2.com/research/paper/68fa7bc81e209131cea8a752.md>)

Recorded evaluation configuration: Not recorded. This record does not assert independent reproduction.

Page 1 of 1. This is an index page, not the full collection.