jhtwosuin/sd15-sg-fullft-75k

SD1.5 full-finetune: guidance-free (GFT) + self-guidance (SG), 75k steps — EMA UNet

Full-finetuned (no LoRA) Stable Diffusion 1.5 UNet that internalizes two guidance signals in a single forward pass:

• rm / gft (guidance-free): CFG scale distilled into the model. Eval knob: cfg (beta = 1/cfg).
• sg (self-guidance): patch-perturbed self-prediction mixing. Eval knob: alpha.

Both are injected through a doubled time_embedding.cond_proj (time_cond_proj_dim = 2 x 512 = 1024): timestep_cond = concat([gse(1/beta - 1), gse(1/alpha - 1)]).

These are the EMA weights (ema_decay 0.9999) from checkpoint-75000.

Training

item	value
init	sd-legacy/stable-diffusion-v1-5 UNet (cond_proj zero-init)
data	LAION Aesthetics 5+ subset, 19.1M images, WebDataset
steps	75,000 (fp16, global batch 256, lr 1e-5 constant + 1k warmup)
objective	dual-guidance distillation (--rm_guidance --sg_guidance), min_alpha 0.25, mixup_ratio 0.2
alpha range	trained on alpha ∈ [0.25, 1.0] — evaluating below 0.25 is extrapolation

Usage

The stock StableDiffusionPipeline cannot express the dual (rm, sg) conditioning — use the custom denoise loop (sample_sg_fullft.py::sg_generate in the training repo), which does a single UNet forward per step (no CFG doubling):

from diffusers import UNet2DConditionModel
unet = UNet2DConditionModel.from_pretrained("jhtwosuin/sd15-sg-fullft-75k")
# vae / text_encoder / tokenizer / scheduler come frozen from sd-legacy/stable-diffusion-v1-5
# timestep_cond = cat([gse(cfg-1), gse(1/alpha-1)]) ; recommended cfg 7.5, alpha ∈ {0.8, 0.5, 0.25}

Non-EMA weights and full training state (optimizer/EMA shadow) exist separately; ask the author if you need them for training resumption.