Human qualitative assessment shows SDXL significantly outperforms SD-v3.5-Medium after fine-tuning, failing on 84 versus 250 of 300 prompts¶
Claim · All claims
Evidence · 1 study · designr2·q2
Currency · newest study 2026 · not yet checked against its sources · not reviewed by a person · recheck by 2029: tool and model accuracy changes quickly
Subclaims¶
Reading the evidence codes
q is the design tier, i the impact, n the sample, and the kind with its r
the rigour judged within that kind: what each code means.
This claim: no standardized effect is recorded in observations/ for it yet, so no benchmark reading is possible.
q2 i? In human-judged comparison against real images, SDXL failed to generate a reasonably close image in 84 of 300 instances while SD-v3.5-Medium failed in 250, so SDXL performs significantly better despite comparable CMMD/KID scores. → Mohammed I. Radaideh 2026
Evidence¶
Mohammed I. Radaideh 2026¶
Mohammed I. Radaideh, Jeremy Moon, Andre Gala-Garza, Emma Son, Yug Shah, Majdi I. Radaideh. (2026). NuclearDiffusion: Text-to-Image Foundation Models for Learning Nuclear Energy Concepts. Preprint submitted to Peer Review. https://arxiv.org/abs/2608.04030 doi:10.48550/arxiv.2608.04030
q2 · i? · design · r2
Human qualitative assessment (subsection 4.2): a human judge picked the checkpoint image closest to the real image across all 300 test prompts; Flux.1 failed all 300 and was excluded. Top SDXL checkpoints scored 50, 48, and 44 points; the "significantly better performance" of SDXL is reported without an effect size.
"the bottom panel of Figure 5 shows that SD-v3.5-Medium failed to generate images reasonably close to the real ones in 250 instances, while SDXL failed in 84 instances. Accordingly, SDXL exhibits significantly better performance than SD-v3.5-Medium."
Discussion¶
Related Claims¶
- The CLIP text-image alignment score failed to capture fine-tuning improvements, contradicting KID, CMMD, and human judgment — related
- Commercial text-to-image models produce accurate images for broad nuclear concepts but fail on specialized engineering prompts — related
- Fine-tuning on the 1,000-image nuclear dataset improves image fidelity of SD models, with KID and CMMD decreasing over training steps — related
- LoRA fine-tuning of Flux.1 failed to learn nuclear concepts, producing wrong-concept images on all 300 test prompts — related
- Generating multiple images per prompt and re-prompting with word variations did not yield accurate nuclear images for failed concepts — related
Page metadata
| Field | Value |
|---|---|
type |
claim |
id |
sdxl-outperforms-sd35-medium-qualitatively |
title |
Human qualitative assessment shows SDXL significantly outperforms SD-v3.5-Medium after fine-tuning, failing on 84 versus 250 of 300 prompts |
description |
Human qualitative assessment shows SDXL significantly outperforms SD-v3.5-Medium after fine-tuning, failing on 84 versus 250 of 300 prompts |
status |
review |
evidence_strength |
moderate |
newest_evidence |
2026 |
recheck_by |
2029 |
generated |
by process:wiki-ingest on 2026-10-11 |
trust |
unverified — no one has confirmed this page's content |
sources |
1 — mohammed-i-radaideh-2026 |
Frontmatter is the page's machine-readable record — see the Schema & Guide.