4-step distill: download BF16, convert to Q5_K_M GGUF at build time

Dockerfile

@@ -39,9 +39,14 @@ RUN chmod +x /app/sd-cli
 
 RUN mkdir -p /app/models
 
-# Download Z-Anime distill 8-step Q5_0 GGUF (~4.51GB)
-RUN curl -fL --retry 3 --retry-delay 5 -o /app/models/z-anime-8steps-q5_0.gguf \
-    "https://huggingface.co/DaNS2025/Z-Anime_8-steps.GGUF/resolve/main/Z-Anime-8steps.q5_0.gguf"
+# Download Z-Anime distill 4-step BF16 (~12.3GB), convert to Q5_K_M GGUF, delete source
+RUN curl -fL --retry 3 --retry-delay 5 -o /app/models/z-anime-distill-4step-bf16.safetensors \
+    "https://huggingface.co/SeeSee21/Z-Anime/resolve/main/diffusion_models/z-anime-distill-4step-bf16.safetensors" \
+    && /app/sd-cli -M convert \
+        -m /app/models/z-anime-distill-4step-bf16.safetensors \
+        -o /app/models/z-anime-distill-4step-Q5_K_M.gguf \
+        --type q5_k_m \
+    && rm /app/models/z-anime-distill-4step-bf16.safetensors
 
 # Download Qwen3-4B text encoder Q8_0 GGUF (~4.28GB)
 RUN curl -fL --retry 3 --retry-delay 5 -o /app/models/qwen3_4b_q8_0.gguf \

README.md

@@ -20,19 +20,21 @@ startup_duration_timeout: 1h
 
 # Z-Anime Image Generation (CPU)
 
-Generate anime images with [Z-Anime 6B](https://huggingface.co/SeeSee21/Z-Anime) (S3-DiT, distill 8-step) via [stable-diffusion.cpp](https://github.com/leejet/stable-diffusion.cpp). Runs on free CPU Spaces.
+Generate anime images with [Z-Anime 6B](https://huggingface.co/SeeSee21/Z-Anime) (S3-DiT, distill 4-step) via [stable-diffusion.cpp](https://github.com/leejet/stable-diffusion.cpp). Runs on free CPU Spaces.
 
 ## Models
 
 | Component | File | Size |
 |-----------|------|------|
-| Diffusion (DiT) | Z-Anime-8steps Q5_0 GGUF | 4.51 GB |
+| Diffusion (DiT) | Z-Anime distill-4step Q5_K_M GGUF | ~4.5 GB |
 | Text Encoder | Qwen3-4B Q8_0 GGUF | 4.28 GB |
 | VAE | ae.safetensors | 168 MB |
 
+GGUF converted at build time from [Z-Anime BF16](https://huggingface.co/SeeSee21/Z-Anime/blob/main/diffusion_models/z-anime-distill-4step-bf16.safetensors).
+
 ## Settings
 
-- **Steps:** 8 (distilled)
+- **Steps:** 4 (distilled)
 - **CFG:** 1.0
 - **Sampler:** euler_ancestral
 - **Resolution:** 512x512 (recommended on CPU)

app.py

@@ -7,12 +7,12 @@ import gradio as gr
 # ---------------------------------------------------------------------------
 # Model paths (downloaded at build time)
 # ---------------------------------------------------------------------------
-DIFFUSION = "/app/models/z-anime-8steps-q5_0.gguf"
+DIFFUSION = "/app/models/z-anime-distill-4step-Q5_K_M.gguf"
 LLM = "/app/models/qwen3_4b_q8_0.gguf"
 VAE = "/app/models/ae.safetensors"
 
 RESOLUTIONS = ["512x512", "768x512", "512x768"]
-STEPS = 8
+STEPS = 4
 CFG = 1.0
 TIMEOUT = 10800
 
@@ -108,8 +108,8 @@ def generate(prompt, negative_prompt, resolution, seed):
 # ---------------------------------------------------------------------------
 with gr.Blocks(title="Z-Anime (CPU)") as demo:
     gr.Markdown(
-        "**[Z-Anime 6B](https://huggingface.co/SeeSee21/Z-Anime)** S3-DiT Q5_0 GGUF "
-        "(distill 8-step) via [sd.cpp](https://github.com/leejet/stable-diffusion.cpp) | "
+        "**[Z-Anime 6B](https://huggingface.co/SeeSee21/Z-Anime)** S3-DiT Q5_K_M GGUF "
+        "(distill 4-step) via [sd.cpp](https://github.com/leejet/stable-diffusion.cpp) | "
         "Free CPU inference"
     )
     with gr.Row():
@@ -122,7 +122,7 @@ with gr.Blocks(title="Z-Anime (CPU)") as demo:
                 res_input = gr.Dropdown(choices=RESOLUTIONS, value="512x512",
                     label="Resolution")
                 seed_input = gr.Number(value=-1, label="Seed (-1=random)", precision=0)
-            gen_btn = gr.Button("Generate (8 steps, CFG 1)", variant="primary", size="lg")
+            gen_btn = gr.Button("Generate (4 steps, CFG 1)", variant="primary", size="lg")
         with gr.Column():
             output_img = gr.Image(type="pil", label="Output")
             status_box = gr.Textbox(label="Status", interactive=False)