Questions on training/evaluation prompt corpora and reward composition
We are reproducing this work for use as a baseline and have found that, for two of the three backbones, the prompt corpora and reward compositions described in Sec. 4.1 do not correspond to those in the released configuration files. We suspect a version difference between the submitted experiments and the current release. Below we state the discrepancies per model; an indication of which configuration file corresponds to which table would be helpful and would likely resolve several of these questions at once.
A. SD3-M (Table 1)
| |
Manuscript (Sec. 4.1) |
Released configs/{cold_start,rtdmd}/sd3m.yaml |
| Stage-1 prompt corpus |
not stated |
dataset/pickscore |
| Stage-2 prompt corpus |
t2i-2M [13] |
dataset/geneval |
| Stage-2 rewards |
HPSv2, CLIPScore |
CLIPScore, HPSv2, PickScore, GenEval (equally weighted) |
B1. As for SD3-M, the prompt corpus used for first-stage distillation is not stated. Could the authors specify it, and confirm whether it is the same corpus used for SD3-M and SD3.5-M?
B2. Section 4.1 specifies the four rewards used for FLUX.2 4B but does not state the prompt corpus on which reward optimization was performed. Could the authors specify it? We note that the released configuration uses the GenEval prompt set, and that several other quantities in flux2_4b.yaml do correspond exactly to the manuscript — the four rewards match, and sample_batch_size=6 with K=24 over the sixteen GPUs stated in Appendix J yields precisely the 64 groups of size 24 reported in Sec. 4.1. If the GenEval prompt set is indeed what was used, it would be helpful to state this in the paper, as it bears on the interpretation of Table 2: the reported GenEval gain (+0.13 over the 4-NFE baseline) would then be partly in-distribution, whereas GenEval2 (+0.03) and OCR (+0.05) would be out-of-distribution, which is consistent with the pattern observed in that table.
B3. Could the authors confirm that the reward weighting is uniform (1:1:1:1)? Appendix J states that "each reward is weighted equally," which matches the released configuration. We ask because GenEval is non-differentiable and therefore contributes through the SubGRPO term but not through the direct-backpropagation term of Eq. (15), so the effective reward composition differs between the two components of the hybrid policy gradient. Is this asymmetry intended?
C. Evaluation prompt sets (both models)
C1. Section 4.1 states that SD3-M is evaluated on prompts sampled from ShareGPT-4o-Image [8]. Could the authors report the size of the evaluation subset and the sampling procedure, or release the prompt list? Since the baselines in Table 1 are partly reproduced by the authors and partly taken from prior work, the evaluation prompt set is required for the comparison to be reproducible.
Questions on training/evaluation prompt corpora and reward composition
We are reproducing this work for use as a baseline and have found that, for two of the three backbones, the prompt corpora and reward compositions described in Sec. 4.1 do not correspond to those in the released configuration files. We suspect a version difference between the submitted experiments and the current release. Below we state the discrepancies per model; an indication of which configuration file corresponds to which table would be helpful and would likely resolve several of these questions at once.
A. SD3-M (Table 1)
B1. As for SD3-M, the prompt corpus used for first-stage distillation is not stated. Could the authors specify it, and confirm whether it is the same corpus used for SD3-M and SD3.5-M?
B2. Section 4.1 specifies the four rewards used for FLUX.2 4B but does not state the prompt corpus on which reward optimization was performed. Could the authors specify it? We note that the released configuration uses the GenEval prompt set, and that several other quantities in
flux2_4b.yamldo correspond exactly to the manuscript — the four rewards match, andsample_batch_size=6withK=24over the sixteen GPUs stated in Appendix J yields precisely the 64 groups of size 24 reported in Sec. 4.1. If the GenEval prompt set is indeed what was used, it would be helpful to state this in the paper, as it bears on the interpretation of Table 2: the reported GenEval gain (+0.13 over the 4-NFE baseline) would then be partly in-distribution, whereas GenEval2 (+0.03) and OCR (+0.05) would be out-of-distribution, which is consistent with the pattern observed in that table.B3. Could the authors confirm that the reward weighting is uniform (1:1:1:1)? Appendix J states that "each reward is weighted equally," which matches the released configuration. We ask because GenEval is non-differentiable and therefore contributes through the SubGRPO term but not through the direct-backpropagation term of Eq. (15), so the effective reward composition differs between the two components of the hybrid policy gradient. Is this asymmetry intended?
C. Evaluation prompt sets (both models)
C1. Section 4.1 states that SD3-M is evaluated on prompts sampled from ShareGPT-4o-Image [8]. Could the authors report the size of the evaluation subset and the sampling procedure, or release the prompt list? Since the baselines in Table 1 are partly reproduced by the authors and partly taken from prior work, the evaluation prompt set is required for the comparison to be reproducible.