Add five-step PCM/CDM checkpoint 1000 and restore model card

Student adapter from qwen21-pcm8-cdm-original-targets-20260923, S3 checkpoint 0001000. Restore supplied README with five-step crossover sigma corrected to 0.94. Preserve both eight-step adapters.

README.md

@@ -33,7 +33,7 @@ tags:
 <h1 style="color: #9334E9;">⚡ Qwen-Image-2.1-P-Turbo</h1>
 <h2>Few-step LoRA adapters for Qwen-Image-2.1</h2>
 <h3>
-  <span style="color: #9334E9;">8 steps</span>
+  <span style="color: #9334E9;">5 or 8 steps</span>
   &nbsp;·&nbsp;
   <span style="color: #9334E9;">No CFG</span>
   &nbsp;·&nbsp;
@@ -43,8 +43,9 @@ tags:
 
 **Qwen-Image-2.1-P-Turbo is a set of LoRA adapters that let
 [`Qwen/Qwen-Image-2.1`](https://huggingface.co/Qwen/Qwen-Image-2.1) generate
-and edit images in only 8 steps. The adapters load on top of the base
-model, so the pipeline, text encoder and VAE stay unchanged. Improved using Qwen.**
+and edit images in only 5 or 8 steps. The adapters load on top of the base
+model, so the pipeline, text encoder and VAE stay unchanged. Training is based
+on DMD. Improved using Qwen.**
 
 > **v0.1: first release, work in progress.** P-Turbo does not yet match the
 > visual quality of the base model. We are still improving the distillation and
@@ -52,12 +53,14 @@ model, so the pipeline, text encoder and VAE stay unchanged. Improved using Qwen
 
 ## Variants
 
-The v0.1 adapter is available in two formats. Load only one at a time.
+Both adapters are v0.1. Pick one based on whether you need quality or speed.
 
-| File | Steps | Format |
+| File | Steps | Trade-off |
 |---|:---:|---|
-| `p_qwen_image_2.1_8step_v0.1.safetensors` | 8 | Diffusers |
-| `p_qwen_image_2.1_8step_v0.1_comfyui_bf16.safetensors` | 8 | ComfyUI |
+| `p_qwen_image_2.1_8step_v0.1.safetensors` | 8 | **Higher quality.** Recommended as the default. |
+| `p_qwen_image_2.1_5step_v0.1.safetensors` | 5 | **Higher speed**, with noticeably lower visual quality. |
+
+Each adapter is trained for its own sigma schedule. Load only one at a time.
 
 ## Quickstart
 
@@ -74,10 +77,13 @@ import torch
 from PIL import Image
 from diffusers import FlowMatchEulerDiscreteScheduler, QwenImage21Pipeline
 
-STEPS = 8
+STEPS = 8  # 8 for higher quality, 5 for higher speed
 
 # The terminal sigma 0 is appended by the scheduler.
-SIGMAS = [1.0, 14 / 15, 6 / 7, 10 / 13, 2 / 3, 6 / 11, 0.4, 2 / 9]
+SIGMAS = {
+    5: [1.0, 0.94, 6 / 7, 2 / 3, 0.4],
+    8: [1.0, 14 / 15, 6 / 7, 10 / 13, 2 / 3, 6 / 11, 0.4, 2 / 9],
+}[STEPS]
 
 pipe = QwenImage21Pipeline.from_pretrained(
     "Qwen/Qwen-Image-2.1", torch_dtype=torch.bfloat16
@@ -132,6 +138,7 @@ image.save("edit_example.png")
   at `shift=1.0` with dynamic shifting off so the sigmas are not shifted twice.
   - 8-step: `1 → 14/15 → 6/7 → 10/13 → 2/3 → 6/11 → 0.4 → 2/9 → 0`
     (shift 2, computed as `σ = 2t / (1 + t)` on evenly spaced `t`)
+  - 5-step: `1 → 0.94 → 6/7 → 2/3 → 0.4 → 0`
 - **No CFG.** Keep `true_cfg_scale=1.0` and do not pass a negative prompt.
 - **Keep the LoRA strength at 1.0.**
 - **Write detailed prompts for text-to-image.** Longer prompts that describe
@@ -141,7 +148,12 @@ image.save("edit_example.png")
 ## Limitations
 
 - This is a first version. Quality is below that of the base model.
+- The 5-step adapter is faster than the 8-step one, but its images are
+  visibly worse.
 - Short or vague text-to-image prompts give weaker results.
+- Trained only at 1K resolution (around 1024 × 1024 pixels). Other resolutions
+  may give worse results. The next version will also be trained at higher
+  resolutions.
 - The adapters are only intended for the settings above. Other step counts,
   schedules or CFG values are not supported.
 

p_qwen_image_2.1_5step_v0.1.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:284cdf6f18983a89035016a39c8b5b74f71edd46b2af46ba801b093e3b09be2b
+size 335606144