Text-to-Image
PEFT
Safetensors
lora

CANDOR: Contrastive Self-Denoising Rewards for Label-Free Text-to-Image Post-Training. Junhyeok Lee, Byung-Hoon Kim, Kyu Sung Choi. Seoul National University, EverEx.

arXiv 2611.25457 Project page Code on GitHub

A label-free reward for Flow-GRPO that scores how much the prompt helps a text-to-image model denoise its own sample, not how well the model denoises it.

CANDOR adapters

LoRA adapters (rank 32, alpha 64) of Stable Diffusion 3.5 Medium post-trained with CANDOR, from CANDOR: Contrastive Self-Denoising Rewards for Label-Free Text-to-Image Post-Training (Junhyeok Lee, Byung-Hoon Kim, Kyu Sung Choi).

Powered by Stability AI. The adapters are derived from Stable Diffusion 3.5 Medium and released under the Stability AI Community License (LICENSE.md); see NOTICE.

Five test prompts from the base model (top) and after CANDOR post-training (bottom): an elephant below a surfboard, two trains, and the texts Abandon All Hope, Bestsellers of 2023 and No Liquids Allowed.

Base model before and after CANDOR post-training. GenEval (left two) and OCR text-rendering (right three) test prompts with shared initial noise; OCR labels: text to render. Hand-picked to show differences.

Highlights

0.60 / 0.63 0.63 → 0.70 1.12 → 1.28 0 labels
Concordance with correctness within same-prompt groups, GenEval / OCR. Absolute self-confidence: 0.48 / 0.51 (chance 0.5). OCR score of SD3.5 Medium with CANDOR as the only reward. ImageReward of the public Flow-GRPO (GenEval) adapter after post-training with CANDOR. No annotations and no reward model: classifier-free guidance already has the empty-prompt branch.

Method

Overview of CANDOR: rollout and re-noising with antithetic probes, the contrastive denoising log-ratio of empty-prompt to prompt errors, and GRPO advantages.

Within the same-prompt groups of GRPO, absolute self-confidence, how well the model denoises its own sample, is dominated by image simplicity. CANDOR rewards the contrast between the unconditional and conditional denoising errors on the same re-noised latent and noise:

D=log⁡MSE(∅)−log⁡MSE(c) D = \log \mathrm{MSE}(\varnothing) - \log \mathrm{MSE}(c)

The log-ratio cancels any difficulty factor that scales both errors, so DD is large when the prompt explains the sample.

Adapters

Folder Trained from
candor_sd35m Stable Diffusion 3.5 Medium
candor_on_flowgrpo_geneval jieliu/SD3.5M-FlowGRPO-GenEval (revision f821fec9)
candor_on_flowgrpo_text jieliu/SD3.5M-FlowGRPO-Text (revision 1a0e6426)
candor_on_flowgrpo_pickscore jieliu/SD3.5M-FlowGRPO-PickScore (revision 10c56697)

Usage

import torch
from diffusers import StableDiffusion3Pipeline
from peft import PeftModel

pipe = StableDiffusion3Pipeline.from_pretrained("stabilityai/stable-diffusion-3.5-medium", torch_dtype=torch.bfloat16)
pipe.transformer = PeftModel.from_pretrained(pipe.transformer, "jhlee0619/candor", subfolder="candor_sd35m").merge_and_unload()
image = pipe("a photo of two trains", height=512, width=512, num_inference_steps=40, guidance_scale=7.0).images[0]

For a candor_on_flowgrpo_* adapter, merge its Flow-GRPO adapter first in the same way.

Results

Task-specific, text–image alignment and aesthetics

Model GenEval OCR CLIPScore Aesthetic
SD3.5 Medium 0.69 0.63 0.284 5.38
+ CANDOR 0.71 0.70 0.285 5.45
Flow-GRPO (GenEval) 0.94 0.71 0.294 5.33
+ CANDOR 0.92 0.73 0.301 5.52
Flow-GRPO (OCR) 0.70 0.92 0.290 5.31
+ CANDOR 0.74 0.92 0.293 5.24
Flow-GRPO (PickScore) 0.75 0.73 0.280 5.91
+ CANDOR 0.74 0.76 0.281 5.97

Human preference

Model PickScore HPSv2.1 ImageReward UnifiedReward
SD3.5 Medium 22.37 0.283 0.91 3.11
+ CANDOR 22.56 0.297 1.09 3.20
Flow-GRPO (GenEval) 22.53 0.280 1.12 3.27
+ CANDOR 22.73 0.305 1.28 3.20
Flow-GRPO (OCR) 22.45 0.284 0.99 3.16
+ CANDOR 22.55 0.293 1.13 3.11
Flow-GRPO (PickScore) 23.47 0.321 1.31 3.41
+ CANDOR 23.55 0.322 1.36 3.43

From the post-training results table of the paper, measured by the authors at 512 × 512, 40 steps and CFG 7.0. GenEval: official task-averaged; OCR: Flow-GRPO test set; other metrics: DrawBench. Each + CANDOR row is the adapter post-trained from the row above it: candor_sd35m, candor_on_flowgrpo_geneval, candor_on_flowgrpo_text and candor_on_flowgrpo_pickscore, from top; Flow-GRPO (OCR) is jieliu/SD3.5M-FlowGRPO-Text. Higher is better for all metrics.

Citation

@article{lee2026candor,
  title   = {{CANDOR}: Contrastive Self-Denoising Rewards for Label-Free Text-to-Image Post-Training},
  author  = {Lee, Junhyeok and Kim, Byung-Hoon and Choi, Kyu Sung},
  journal = {arXiv preprint arXiv:2611.25457},
  year    = {2026}
}
Downloads last month
-
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support

Model tree for jhlee0619/candor

Adapter
(124)
this model

Paper for jhlee0619/candor