SDXS-1B (CLIP version)

Text-to-image diffusion model on a custom UNet with CLIP text encoder (converted from Qwen).

Architecture

• UNet: ~1.585B params, cross_attention_dim=768
  • block_out_channels: [320, 640, 1280, 1536]
  • layers_per_block: [4, 3, 2, 2]
  • transformer_layers_per_block: [2, 2, 3, 4]
• Text encoder: CLIPTextModel (768 hidden, 12 layers, 248 max tokens) — LongCLIP-style
• VAE: AsymmetricAutoencoderKL (32ch, 8x encoder / 16x decoder, built-in 2x upscaler)
• Scheduler: FlowMatchEulerDiscreteScheduler, shift=5.0, 40 steps default
• Embeddings: hidden_states[-2] + final_layer_norm

Usage

import torch
from pipeline_sdxs import SdxsPipeline

pipe = SdxsPipeline.from_pretrained(
    "recoilme/sdxs",
    torch_dtype=torch.float16,
    trust_remote_code=True,
).to("cuda")

image = pipe(
    prompt="1girl, smiling, red eyes, blue hair, white shirt",
    negative_prompt="bad quality grainy image with low details, incomplete text, despite numerous technical flaws and distorted figures",
    width=1024,
    height=1280,
    guidance_scale=5.0,
    seed=0,
)[0][0]

image.show()

Notes

• Converted from the Qwen-based SDXS-1B: text encoder replaced with CLIP, UNet cross-attention migrated 2048 → 768 (warm-started projections).
• No refine_prompts (CLIP is not an LLM) and no image conditioning.
• Training: flow matching (velocity), cfg-dropout 0.10, timestep shift aligned with the scheduler.