Upload eval_report/lora_model_ablation_report.html with huggingface_hub

Browse files

Files changed (1) hide show

eval_report/lora_model_ablation_report.html +103 -46

eval_report/lora_model_ablation_report.html CHANGED Viewed

@@ -38,11 +38,11 @@ h3, h4 { letter-spacing: 0; }
 .panel { padding: 20px; }
 .panel p { margin: 0 0 12px; }
 .panel p:last-child { margin-bottom: 0; }
-.metrics { display: grid; grid-template-columns: repeat(3, minmax(0, 1fr)); gap: 10px; }
 .metric { padding: 14px; background: var(--soft); border-radius: 6px; min-width: 0; }
 .metric strong { display: block; font-size: 22px; }
 .metric span { color: var(--muted); font-size: 12px; }
-.model-grid { display: grid; grid-template-columns: repeat(3, minmax(0, 1fr)); gap: 14px; }
 .model-card { padding: 18px; min-width: 0; }
 .model-title { display: flex; justify-content: space-between; gap: 12px; align-items: baseline; }
 .model-title h3 { margin: 0; font-size: 18px; }
@@ -67,21 +67,31 @@ figcaption span { color: var(--muted); }
 .eval-heading { display: flex; justify-content: space-between; gap: 14px; align-items: baseline; margin-bottom: 10px; }
 .eval-heading h3 { margin: 0; font-size: 18px; }
 .eval-heading span { color: var(--muted); font-size: 12px; }
-.result-grid { display: grid; grid-template-columns: repeat(4, minmax(0, 1fr)); gap: 12px; }
 .result-cell { overflow: hidden; min-width: 0; }
 .result-cell a { display: block; aspect-ratio: 1 / 1; background: #eceae2; }
 .source-cell a { background: #e7ebe7; }
 .result-cell figcaption { display: grid; gap: 2px; min-height: 56px; padding: 9px 10px 11px; }
 .result-cell figcaption strong { font-size: 13px; }
 .result-cell figcaption span { font-size: 12px; overflow-wrap: anywhere; }
-.findings { display: grid; grid-template-columns: repeat(3, minmax(0, 1fr)); gap: 12px; }
 .finding { padding: 16px; background: var(--soft); border-radius: 8px; border: 1px solid var(--line); }
 .finding h3 { margin: 0 0 6px; font-size: 16px; }
 .finding p { margin: 0; color: var(--muted); }
 footer { padding: 18px 0 32px; border-top: 1px solid var(--line); color: var(--muted); }
 @media (max-width: 1100px) {
-  .overview, .model-grid, .training-grid, .findings { grid-template-columns: 1fr; }
   .result-grid { grid-template-columns: repeat(2, minmax(0, 1fr)); }
 }
 @media (max-width: 620px) {
   header, main, footer { width: min(100% - 24px, 1480px); }
@@ -102,7 +112,7 @@ footer { padding: 18px 0 32px; border-top: 1px solid var(--line); color: var(--m
 <body>
   <header>
     <p class="kicker">LoRA training progress report</p>
-    <h1>Comparing style-transfer LoRA edits across Flux Klein, SD 1.5, and SD 3.5</h1>
     <p class="subtitle">This report summarizes the ablation in <code>ABLATION.md</code> and visualizes the generated artifacts under <code>outputs/</code>. The goal is to learn a compact LoRA that turns a normal source image into the project-specific stylized target look while preserving the source image layout and recognizable structure.</p>
   </header>
   <main>
@@ -110,14 +120,15 @@ footer { padding: 18px 0 32px; border-top: 1px solid var(--line); color: var(--m
       <article class="panel">
         <h2>Purpose</h2>
         <p>The training data is paired: each conditioning image provides the source structure, and each target image provides the desired rendering style. A successful model should transfer the luminous, graphic label style from the targets without drifting away from the source image content.</p>
-        <p>The evaluation set uses five held-out source images. Each result row below starts with the eval source image on the left, followed by the three model outputs for direct structure and style comparison.</p>
       </article>
       <article class="panel">
         <h2>Progress So Far</h2>
         <div class="metrics">
           <div class="metric"><strong>25</strong><span>paired training examples</span></div>
           <div class="metric"><strong>5</strong><span>held-out eval sources</span></div>
-          <div class="metric"><strong>3</strong><span>model families compared</span></div>
         </div>
       </article>
     </section>
@@ -127,9 +138,29 @@ footer { padding: 18px 0 32px; border-top: 1px solid var(--line); color: var(--m
       <div class="model-grid">
       <article class="model-card">
         <div class="model-title">
-          <h3>FLUX.2 Klein 4B</h3>
           <span>Best result so far</span>
         </div>
         <dl>
           <div><dt>Size</dt><dd>4B</dd></div>
           <div><dt>Base repo</dt><dd><code>black-forest-labs/FLUX.2-klein-base-4B</code></dd></div>
@@ -139,10 +170,10 @@ footer { padding: 18px 0 32px; border-top: 1px solid var(--line); color: var(--m
           <div><dt>Conditioning</dt><dd>source_latents_concatenated_to_noisy_target_latents</dd></div>
           <div><dt>Checkpoint</dt><dd><code>../local_edit_lora/flux2_klein_base/checkpoint-002000</code></dd></div>
         </dl>
-        <p>Native multi-reference/editing path. After 2000 LoRA steps it best preserves the eval-source structure while transferring the target stylized label/rendering look.</p>
-        <p class="decision">Recommended path for deployment work: continue with ONNX then CANN export.</p>
       </article>
       <article class="model-card">
         <div class="model-title">
           <h3>Stable Diffusion 1.5</h3>
@@ -160,7 +191,7 @@ footer { padding: 18px 0 32px; border-top: 1px solid var(--line); color: var(--m
         <p>Pix2Pix-style pipeline learns parts of the style but quality is weaker. High image guidance ignores the source image; weak image guidance fails to transfer enough style.</p>
         <p class="decision">Keep as a low-cost baseline; next useful branch is ControlNet plus IP-Adapter with stronger base weights.</p>
       </article>
       <article class="model-card">
         <div class="model-title">
           <h3>Stable Diffusion 3.5 Medium</h3>
@@ -198,7 +229,7 @@ footer { padding: 18px 0 32px; border-top: 1px solid var(--line); color: var(--m
         </div>
         <h4>Pair 001</h4>
       </article>
       <article class="pair-card">
         <div class="pair-images">
           <figure>
@@ -212,7 +243,7 @@ footer { padding: 18px 0 32px; border-top: 1px solid var(--line); color: var(--m
         </div>
         <h4>Pair 002</h4>
       </article>
       <article class="pair-card">
         <div class="pair-images">
           <figure>
@@ -226,7 +257,7 @@ footer { padding: 18px 0 32px; border-top: 1px solid var(--line); color: var(--m
         </div>
         <h4>Pair 004</h4>
       </article>
       <article class="pair-card">
         <div class="pair-images">
           <figure>
@@ -240,7 +271,7 @@ footer { padding: 18px 0 32px; border-top: 1px solid var(--line); color: var(--m
         </div>
         <h4>Pair 009</h4>
       </article>
       <article class="pair-card">
         <div class="pair-images">
           <figure>
@@ -254,7 +285,7 @@ footer { padding: 18px 0 32px; border-top: 1px solid var(--line); color: var(--m
         </div>
         <h4>Pair 017</h4>
       </article>
       <article class="pair-card">
         <div class="pair-images">
           <figure>
@@ -274,7 +305,8 @@ footer { padding: 18px 0 32px; border-top: 1px solid var(--line); color: var(--m
     <section>
       <h2>Qualitative Findings</h2>
       <div class="findings">
-        <article class="finding"><h3>Flux Klein</h3><p>Best current direction: learns the style and remains closest to the editable-source objective after 2000 steps.</p></article>
         <article class="finding"><h3>SD 1.5</h3><p>Small and cheap, but current pix2pix-style behavior is weak and sensitive to image-guidance tradeoffs.</p></article>
         <article class="finding"><h3>SD 3.5</h3><p>Learns style, but source-structure preservation is not reliable enough in this setup, so it is on pause.</p></article>
       </div>
@@ -282,7 +314,7 @@ footer { padding: 18px 0 32px; border-top: 1px solid var(--line); color: var(--m
     <section>
       <h2>Held-Out Eval Comparison</h2>
       <section class="eval-row">
         <div class="eval-heading">
           <h3>Eval 1: eval01.jpg</h3>
@@ -293,25 +325,30 @@ footer { padding: 18px 0 32px; border-top: 1px solid var(--line); color: var(--m
             <a href="lora_model_ablation_report_images/eval_eval01.jpg"><img src="lora_model_ablation_report_images/eval_eval01.jpg" alt="Eval source eval01.jpg"></a>
             <figcaption><strong>Eval source</strong><span>1000x667 · 0.32 MB</span></figcaption>
           </figure>
           <figure class="result-cell">
             <a href="lora_model_ablation_report_images/flux2_klein_base_eval01-001.png"><img src="lora_model_ablation_report_images/flux2_klein_base_eval01-001.png" alt="Flux Klein output for eval01.jpg"></a>
             <figcaption><strong>Flux Klein</strong><span>768x768 · 0.44 MB</span></figcaption>
           </figure>
           <figure class="result-cell">
             <a href="lora_model_ablation_report_images/sd15_eval01-001.png"><img src="lora_model_ablation_report_images/sd15_eval01-001.png" alt="SD 1.5 output for eval01.jpg"></a>
             <figcaption><strong>SD 1.5</strong><span>512x512 · 0.28 MB</span></figcaption>
           </figure>
           <figure class="result-cell">
             <a href="lora_model_ablation_report_images/sd35_medium_eval01-001.png"><img src="lora_model_ablation_report_images/sd35_medium_eval01-001.png" alt="SD 3.5 output for eval01.jpg"></a>
             <figcaption><strong>SD 3.5</strong><span>1024x1024 · 1.09 MB</span></figcaption>
           </figure>
         </div>
       </section>
       <section class="eval-row">
         <div class="eval-heading">
           <h3>Eval 2: eval02.jpg</h3>
@@ -322,25 +359,30 @@ footer { padding: 18px 0 32px; border-top: 1px solid var(--line); color: var(--m
             <a href="lora_model_ablation_report_images/eval_eval02.jpg"><img src="lora_model_ablation_report_images/eval_eval02.jpg" alt="Eval source eval02.jpg"></a>
             <figcaption><strong>Eval source</strong><span>2000x1333 · 0.86 MB</span></figcaption>
           </figure>
           <figure class="result-cell">
             <a href="lora_model_ablation_report_images/flux2_klein_base_eval02-002.png"><img src="lora_model_ablation_report_images/flux2_klein_base_eval02-002.png" alt="Flux Klein output for eval02.jpg"></a>
             <figcaption><strong>Flux Klein</strong><span>768x768 · 0.33 MB</span></figcaption>
           </figure>
           <figure class="result-cell">
             <a href="lora_model_ablation_report_images/sd15_eval02-002.png"><img src="lora_model_ablation_report_images/sd15_eval02-002.png" alt="SD 1.5 output for eval02.jpg"></a>
             <figcaption><strong>SD 1.5</strong><span>512x512 · 0.35 MB</span></figcaption>
           </figure>
           <figure class="result-cell">
             <a href="lora_model_ablation_report_images/sd35_medium_eval02-002.png"><img src="lora_model_ablation_report_images/sd35_medium_eval02-002.png" alt="SD 3.5 output for eval02.jpg"></a>
             <figcaption><strong>SD 3.5</strong><span>1024x1024 · 1.03 MB</span></figcaption>
           </figure>
         </div>
       </section>
       <section class="eval-row">
         <div class="eval-heading">
           <h3>Eval 3: eval03.jpg</h3>
@@ -351,25 +393,30 @@ footer { padding: 18px 0 32px; border-top: 1px solid var(--line); color: var(--m
             <a href="lora_model_ablation_report_images/eval_eval03.jpg"><img src="lora_model_ablation_report_images/eval_eval03.jpg" alt="Eval source eval03.jpg"></a>
             <figcaption><strong>Eval source</strong><span>612x408 · 0.02 MB</span></figcaption>
           </figure>
           <figure class="result-cell">
             <a href="lora_model_ablation_report_images/flux2_klein_base_eval03-003.png"><img src="lora_model_ablation_report_images/flux2_klein_base_eval03-003.png" alt="Flux Klein output for eval03.jpg"></a>
             <figcaption><strong>Flux Klein</strong><span>768x768 · 0.42 MB</span></figcaption>
           </figure>
           <figure class="result-cell">
             <a href="lora_model_ablation_report_images/sd15_eval03-003.png"><img src="lora_model_ablation_report_images/sd15_eval03-003.png" alt="SD 1.5 output for eval03.jpg"></a>
             <figcaption><strong>SD 1.5</strong><span>512x512 · 0.25 MB</span></figcaption>
           </figure>
           <figure class="result-cell">
             <a href="lora_model_ablation_report_images/sd35_medium_eval03-003.png"><img src="lora_model_ablation_report_images/sd35_medium_eval03-003.png" alt="SD 3.5 output for eval03.jpg"></a>
             <figcaption><strong>SD 3.5</strong><span>1024x1024 · 0.85 MB</span></figcaption>
           </figure>
         </div>
       </section>
       <section class="eval-row">
         <div class="eval-heading">
           <h3>Eval 4: eval04.webp</h3>
@@ -380,25 +427,30 @@ footer { padding: 18px 0 32px; border-top: 1px solid var(--line); color: var(--m
             <a href="lora_model_ablation_report_images/eval_eval04.webp"><img src="lora_model_ablation_report_images/eval_eval04.webp" alt="Eval source eval04.webp"></a>
             <figcaption><strong>Eval source</strong><span>2560x1707 · 0.13 MB</span></figcaption>
           </figure>
           <figure class="result-cell">
             <a href="lora_model_ablation_report_images/flux2_klein_base_eval04-004.png"><img src="lora_model_ablation_report_images/flux2_klein_base_eval04-004.png" alt="Flux Klein output for eval04.webp"></a>
             <figcaption><strong>Flux Klein</strong><span>768x768 · 0.45 MB</span></figcaption>
           </figure>
           <figure class="result-cell">
             <a href="lora_model_ablation_report_images/sd15_eval04-004.png"><img src="lora_model_ablation_report_images/sd15_eval04-004.png" alt="SD 1.5 output for eval04.webp"></a>
             <figcaption><strong>SD 1.5</strong><span>512x512 · 0.28 MB</span></figcaption>
           </figure>
           <figure class="result-cell">
             <a href="lora_model_ablation_report_images/sd35_medium_eval04-004.png"><img src="lora_model_ablation_report_images/sd35_medium_eval04-004.png" alt="SD 3.5 output for eval04.webp"></a>
             <figcaption><strong>SD 3.5</strong><span>1024x1024 · 1.10 MB</span></figcaption>
           </figure>
         </div>
       </section>
       <section class="eval-row">
         <div class="eval-heading">
           <h3>Eval 5: eval05.jpg</h3>
@@ -409,29 +461,34 @@ footer { padding: 18px 0 32px; border-top: 1px solid var(--line); color: var(--m
             <a href="lora_model_ablation_report_images/eval_eval05.jpg"><img src="lora_model_ablation_report_images/eval_eval05.jpg" alt="Eval source eval05.jpg"></a>
             <figcaption><strong>Eval source</strong><span>900x584 · 0.13 MB</span></figcaption>
           </figure>
           <figure class="result-cell">
             <a href="lora_model_ablation_report_images/flux2_klein_base_eval05-005.png"><img src="lora_model_ablation_report_images/flux2_klein_base_eval05-005.png" alt="Flux Klein output for eval05.jpg"></a>
             <figcaption><strong>Flux Klein</strong><span>768x768 · 0.39 MB</span></figcaption>
           </figure>
           <figure class="result-cell">
             <a href="lora_model_ablation_report_images/sd15_eval05-005.png"><img src="lora_model_ablation_report_images/sd15_eval05-005.png" alt="SD 1.5 output for eval05.jpg"></a>
             <figcaption><strong>SD 1.5</strong><span>512x512 · 0.26 MB</span></figcaption>
           </figure>
           <figure class="result-cell">
             <a href="lora_model_ablation_report_images/sd35_medium_eval05-005.png"><img src="lora_model_ablation_report_images/sd35_medium_eval05-005.png" alt="SD 3.5 output for eval05.jpg"></a>
             <figcaption><strong>SD 3.5</strong><span>1024x1024 · 0.97 MB</span></figcaption>
           </figure>
         </div>
       </section>
     </section>
   </main>
   <footer>
-    Generated from <code>ABLATION.md</code>, <code>dataset/train</code>, <code>dataset/conditioning</code>, <code>dataset/eval</code>, and <code>outputs/ablation/local_edit_eval</code>. Image paths are repo-relative from this report location.
   </footer>
 </body>
 </html>

 .panel { padding: 20px; }
 .panel p { margin: 0 0 12px; }
 .panel p:last-child { margin-bottom: 0; }
+.metrics { display: grid; grid-template-columns: repeat(4, minmax(0, 1fr)); gap: 10px; }
 .metric { padding: 14px; background: var(--soft); border-radius: 6px; min-width: 0; }
 .metric strong { display: block; font-size: 22px; }
 .metric span { color: var(--muted); font-size: 12px; }
+.model-grid { display: grid; grid-template-columns: repeat(4, minmax(0, 1fr)); gap: 14px; }
 .model-card { padding: 18px; min-width: 0; }
 .model-title { display: flex; justify-content: space-between; gap: 12px; align-items: baseline; }
 .model-title h3 { margin: 0; font-size: 18px; }
 .eval-heading { display: flex; justify-content: space-between; gap: 14px; align-items: baseline; margin-bottom: 10px; }
 .eval-heading h3 { margin: 0; font-size: 18px; }
 .eval-heading span { color: var(--muted); font-size: 12px; }
+.result-grid { display: grid; grid-template-columns: repeat(5, minmax(0, 1fr)); gap: 12px; }
 .result-cell { overflow: hidden; min-width: 0; }
 .result-cell a { display: block; aspect-ratio: 1 / 1; background: #eceae2; }
 .source-cell a { background: #e7ebe7; }
 .result-cell figcaption { display: grid; gap: 2px; min-height: 56px; padding: 9px 10px 11px; }
 .result-cell figcaption strong { font-size: 13px; }
 .result-cell figcaption span { font-size: 12px; overflow-wrap: anywhere; }
+.findings { display: grid; grid-template-columns: repeat(4, minmax(0, 1fr)); gap: 12px; }
 .finding { padding: 16px; background: var(--soft); border-radius: 8px; border: 1px solid var(--line); }
 .finding h3 { margin: 0 0 6px; font-size: 16px; }
 .finding p { margin: 0; color: var(--muted); }
 footer { padding: 18px 0 32px; border-top: 1px solid var(--line); color: var(--muted); }
+@media (max-width: 1300px) {
+  .model-grid { grid-template-columns: repeat(2, minmax(0, 1fr)); }
+  .result-grid { grid-template-columns: repeat(3, minmax(0, 1fr)); }
+  .findings { grid-template-columns: repeat(2, minmax(0, 1fr)); }
+}
 @media (max-width: 1100px) {
+  .overview, .training-grid { grid-template-columns: 1fr; }
+  .metrics { grid-template-columns: repeat(2, minmax(0, 1fr)); }
+}
+@media (max-width: 800px) {
   .result-grid { grid-template-columns: repeat(2, minmax(0, 1fr)); }
+  .model-grid { grid-template-columns: 1fr; }
+  .findings { grid-template-columns: 1fr; }
 }
 @media (max-width: 620px) {
   header, main, footer { width: min(100% - 24px, 1480px); }
 <body>
   <header>
     <p class="kicker">LoRA training progress report</p>
+    <h1>Comparing style-transfer LoRA edits across Mage Flow, Flux Klein, SD 1.5, and SD 3.5</h1>
     <p class="subtitle">This report summarizes the ablation in <code>ABLATION.md</code> and visualizes the generated artifacts under <code>outputs/</code>. The goal is to learn a compact LoRA that turns a normal source image into the project-specific stylized target look while preserving the source image layout and recognizable structure.</p>
   </header>
   <main>
       <article class="panel">
         <h2>Purpose</h2>
         <p>The training data is paired: each conditioning image provides the source structure, and each target image provides the desired rendering style. A successful model should transfer the luminous, graphic label style from the targets without drifting away from the source image content.</p>
+        <p>The evaluation set uses five held-out source images. Each result row below starts with the eval source image on the left, followed by the four model outputs for direct structure and style comparison.</p>
       </article>
       <article class="panel">
         <h2>Progress So Far</h2>
         <div class="metrics">
           <div class="metric"><strong>25</strong><span>paired training examples</span></div>
           <div class="metric"><strong>5</strong><span>held-out eval sources</span></div>
+          <div class="metric"><strong>4</strong><span>model families compared</span></div>
+          <div class="metric"><strong>100</strong><span>training epochs (Mage Flow)</span></div>
         </div>
       </article>
     </section>
       <div class="model-grid">
       <article class="model-card">
         <div class="model-title">
+          <h3>Mage Flow Edit Base</h3>
           <span>Best result so far</span>
         </div>
+        <dl>
+          <div><dt>Size</dt><dd>~3B (transformer)</dd></div>
+          <div><dt>Base repo</dt><dd><code>microsoft/Mage-Flow-Edit-Base</code></dd></div>
+          <div><dt>Trainer</dt><dd><code>train_lora.py</code> (custom flow-matching LoRA)</dd></div>
+          <div><dt>Optimizer</dt><dd>Muon (weight_decay=0.01)</dd></div>
+          <div><dt>Epochs</dt><dd>100</dd></div>
+          <div><dt>LoRA</dt><dd>rank 16, alpha 16</dd></div>
+          <div><dt>LR</dt><dd>1e-3 (cosine schedule)</dd></div>
+          <div><dt>Inference</dt><dd>30 steps, CFG 3.0</dd></div>
+          <div><dt>Checkpoint</dt><dd><code>outputs/lora/base_muon/lora_final.pt</code></dd></div>
+        </dl>
+        <p>Native flow-matching editing path with Muon optimizer. After 100 LoRA epochs it produces the best style transfer results — strong luminous aura rendering while faithfully preserving the source image structure and identity.</p>
+        <p class="decision">Recommended path for deployment: Mage Flow delivers the strongest style-transfer fidelity among all models tested.</p>
+      </article>
+      <article class="model-card">
+        <div class="model-title">
+          <h3>FLUX.2 Klein 4B</h3>
+          <span>Strong baseline</span>
+        </div>
         <dl>
           <div><dt>Size</dt><dd>4B</dd></div>
           <div><dt>Base repo</dt><dd><code>black-forest-labs/FLUX.2-klein-base-4B</code></dd></div>
           <div><dt>Conditioning</dt><dd>source_latents_concatenated_to_noisy_target_latents</dd></div>
           <div><dt>Checkpoint</dt><dd><code>../local_edit_lora/flux2_klein_base/checkpoint-002000</code></dd></div>
         </dl>
+        <p>Native multi-reference/editing path. After 2000 LoRA steps it preserves the eval-source structure while transferring the target stylized label/rendering look.</p>
+        <p class="decision">Continue with ONNX then CANN export as an alternative deployment path.</p>
       </article>
       <article class="model-card">
         <div class="model-title">
           <h3>Stable Diffusion 1.5</h3>
         <p>Pix2Pix-style pipeline learns parts of the style but quality is weaker. High image guidance ignores the source image; weak image guidance fails to transfer enough style.</p>
         <p class="decision">Keep as a low-cost baseline; next useful branch is ControlNet plus IP-Adapter with stronger base weights.</p>
       </article>
       <article class="model-card">
         <div class="model-title">
           <h3>Stable Diffusion 3.5 Medium</h3>
         </div>
         <h4>Pair 001</h4>
       </article>
       <article class="pair-card">
         <div class="pair-images">
           <figure>
         </div>
         <h4>Pair 002</h4>
       </article>
       <article class="pair-card">
         <div class="pair-images">
           <figure>
         </div>
         <h4>Pair 004</h4>
       </article>
       <article class="pair-card">
         <div class="pair-images">
           <figure>
         </div>
         <h4>Pair 009</h4>
       </article>
       <article class="pair-card">
         <div class="pair-images">
           <figure>
         </div>
         <h4>Pair 017</h4>
       </article>
       <article class="pair-card">
         <div class="pair-images">
           <figure>
     <section>
       <h2>Qualitative Findings</h2>
       <div class="findings">
+        <article class="finding"><h3>Mage Flow</h3><p>Best overall result. Muon-optimized LoRA (rank 16, 100 epochs) produces the cleanest aura-style transfer — strong luminous glow, excellent source-structure preservation, and cinematic color rendering. The flow-matching editing path is a natural fit for this paired style-transfer task.</p></article>
+        <article class="finding"><h3>Flux Klein</h3><p>Strong baseline: learns the style and remains close to the editable-source objective after 2000 steps. Slightly less vivid color transfer than Mage Flow.</p></article>
         <article class="finding"><h3>SD 1.5</h3><p>Small and cheap, but current pix2pix-style behavior is weak and sensitive to image-guidance tradeoffs.</p></article>
         <article class="finding"><h3>SD 3.5</h3><p>Learns style, but source-structure preservation is not reliable enough in this setup, so it is on pause.</p></article>
       </div>
     <section>
       <h2>Held-Out Eval Comparison</h2>
       <section class="eval-row">
         <div class="eval-heading">
           <h3>Eval 1: eval01.jpg</h3>
             <a href="lora_model_ablation_report_images/eval_eval01.jpg"><img src="lora_model_ablation_report_images/eval_eval01.jpg" alt="Eval source eval01.jpg"></a>
             <figcaption><strong>Eval source</strong><span>1000x667 · 0.32 MB</span></figcaption>
           </figure>
+          <figure class="result-cell">
+            <a href="lora_model_ablation_report_images/mage_flow_eval01-001.png"><img src="lora_model_ablation_report_images/mage_flow_eval01-001.png" alt="Mage Flow output for eval01.jpg"></a>
+            <figcaption><strong>Mage Flow</strong><span>768x512 · 0.33 MB</span></figcaption>
+          </figure>
           <figure class="result-cell">
             <a href="lora_model_ablation_report_images/flux2_klein_base_eval01-001.png"><img src="lora_model_ablation_report_images/flux2_klein_base_eval01-001.png" alt="Flux Klein output for eval01.jpg"></a>
             <figcaption><strong>Flux Klein</strong><span>768x768 · 0.44 MB</span></figcaption>
           </figure>
           <figure class="result-cell">
             <a href="lora_model_ablation_report_images/sd15_eval01-001.png"><img src="lora_model_ablation_report_images/sd15_eval01-001.png" alt="SD 1.5 output for eval01.jpg"></a>
             <figcaption><strong>SD 1.5</strong><span>512x512 · 0.28 MB</span></figcaption>
           </figure>
           <figure class="result-cell">
             <a href="lora_model_ablation_report_images/sd35_medium_eval01-001.png"><img src="lora_model_ablation_report_images/sd35_medium_eval01-001.png" alt="SD 3.5 output for eval01.jpg"></a>
             <figcaption><strong>SD 3.5</strong><span>1024x1024 · 1.09 MB</span></figcaption>
           </figure>
         </div>
       </section>
       <section class="eval-row">
         <div class="eval-heading">
           <h3>Eval 2: eval02.jpg</h3>
             <a href="lora_model_ablation_report_images/eval_eval02.jpg"><img src="lora_model_ablation_report_images/eval_eval02.jpg" alt="Eval source eval02.jpg"></a>
             <figcaption><strong>Eval source</strong><span>2000x1333 · 0.86 MB</span></figcaption>
           </figure>
+          <figure class="result-cell">
+            <a href="lora_model_ablation_report_images/mage_flow_eval02-002.png"><img src="lora_model_ablation_report_images/mage_flow_eval02-002.png" alt="Mage Flow output for eval02.jpg"></a>
+            <figcaption><strong>Mage Flow</strong><span>768x512 · 0.32 MB</span></figcaption>
+          </figure>
           <figure class="result-cell">
             <a href="lora_model_ablation_report_images/flux2_klein_base_eval02-002.png"><img src="lora_model_ablation_report_images/flux2_klein_base_eval02-002.png" alt="Flux Klein output for eval02.jpg"></a>
             <figcaption><strong>Flux Klein</strong><span>768x768 · 0.33 MB</span></figcaption>
           </figure>
           <figure class="result-cell">
             <a href="lora_model_ablation_report_images/sd15_eval02-002.png"><img src="lora_model_ablation_report_images/sd15_eval02-002.png" alt="SD 1.5 output for eval02.jpg"></a>
             <figcaption><strong>SD 1.5</strong><span>512x512 · 0.35 MB</span></figcaption>
           </figure>
           <figure class="result-cell">
             <a href="lora_model_ablation_report_images/sd35_medium_eval02-002.png"><img src="lora_model_ablation_report_images/sd35_medium_eval02-002.png" alt="SD 3.5 output for eval02.jpg"></a>
             <figcaption><strong>SD 3.5</strong><span>1024x1024 · 1.03 MB</span></figcaption>
           </figure>
         </div>
       </section>
       <section class="eval-row">
         <div class="eval-heading">
           <h3>Eval 3: eval03.jpg</h3>
             <a href="lora_model_ablation_report_images/eval_eval03.jpg"><img src="lora_model_ablation_report_images/eval_eval03.jpg" alt="Eval source eval03.jpg"></a>
             <figcaption><strong>Eval source</strong><span>612x408 · 0.02 MB</span></figcaption>
           </figure>
+          <figure class="result-cell">
+            <a href="lora_model_ablation_report_images/mage_flow_eval03-003.png"><img src="lora_model_ablation_report_images/mage_flow_eval03-003.png" alt="Mage Flow output for eval03.jpg"></a>
+            <figcaption><strong>Mage Flow</strong><span>768x512 · 0.27 MB</span></figcaption>
+          </figure>
           <figure class="result-cell">
             <a href="lora_model_ablation_report_images/flux2_klein_base_eval03-003.png"><img src="lora_model_ablation_report_images/flux2_klein_base_eval03-003.png" alt="Flux Klein output for eval03.jpg"></a>
             <figcaption><strong>Flux Klein</strong><span>768x768 · 0.42 MB</span></figcaption>
           </figure>
           <figure class="result-cell">
             <a href="lora_model_ablation_report_images/sd15_eval03-003.png"><img src="lora_model_ablation_report_images/sd15_eval03-003.png" alt="SD 1.5 output for eval03.jpg"></a>
             <figcaption><strong>SD 1.5</strong><span>512x512 · 0.25 MB</span></figcaption>
           </figure>
           <figure class="result-cell">
             <a href="lora_model_ablation_report_images/sd35_medium_eval03-003.png"><img src="lora_model_ablation_report_images/sd35_medium_eval03-003.png" alt="SD 3.5 output for eval03.jpg"></a>
             <figcaption><strong>SD 3.5</strong><span>1024x1024 · 0.85 MB</span></figcaption>
           </figure>
         </div>
       </section>
       <section class="eval-row">
         <div class="eval-heading">
           <h3>Eval 4: eval04.webp</h3>
             <a href="lora_model_ablation_report_images/eval_eval04.webp"><img src="lora_model_ablation_report_images/eval_eval04.webp" alt="Eval source eval04.webp"></a>
             <figcaption><strong>Eval source</strong><span>2560x1707 · 0.13 MB</span></figcaption>
           </figure>
+          <figure class="result-cell">
+            <a href="lora_model_ablation_report_images/mage_flow_eval04-004.png"><img src="lora_model_ablation_report_images/mage_flow_eval04-004.png" alt="Mage Flow output for eval04.webp"></a>
+            <figcaption><strong>Mage Flow</strong><span>768x512 · 0.33 MB</span></figcaption>
+          </figure>
           <figure class="result-cell">
             <a href="lora_model_ablation_report_images/flux2_klein_base_eval04-004.png"><img src="lora_model_ablation_report_images/flux2_klein_base_eval04-004.png" alt="Flux Klein output for eval04.webp"></a>
             <figcaption><strong>Flux Klein</strong><span>768x768 · 0.45 MB</span></figcaption>
           </figure>
           <figure class="result-cell">
             <a href="lora_model_ablation_report_images/sd15_eval04-004.png"><img src="lora_model_ablation_report_images/sd15_eval04-004.png" alt="SD 1.5 output for eval04.webp"></a>
             <figcaption><strong>SD 1.5</strong><span>512x512 · 0.28 MB</span></figcaption>
           </figure>
           <figure class="result-cell">
             <a href="lora_model_ablation_report_images/sd35_medium_eval04-004.png"><img src="lora_model_ablation_report_images/sd35_medium_eval04-004.png" alt="SD 3.5 output for eval04.webp"></a>
             <figcaption><strong>SD 3.5</strong><span>1024x1024 · 1.10 MB</span></figcaption>
           </figure>
         </div>
       </section>
       <section class="eval-row">
         <div class="eval-heading">
           <h3>Eval 5: eval05.jpg</h3>
             <a href="lora_model_ablation_report_images/eval_eval05.jpg"><img src="lora_model_ablation_report_images/eval_eval05.jpg" alt="Eval source eval05.jpg"></a>
             <figcaption><strong>Eval source</strong><span>900x584 · 0.13 MB</span></figcaption>
           </figure>
+          <figure class="result-cell">
+            <a href="lora_model_ablation_report_images/mage_flow_eval05-005.png"><img src="lora_model_ablation_report_images/mage_flow_eval05-005.png" alt="Mage Flow output for eval05.jpg"></a>
+            <figcaption><strong>Mage Flow</strong><span>768x496 · 0.33 MB</span></figcaption>
+          </figure>
           <figure class="result-cell">
             <a href="lora_model_ablation_report_images/flux2_klein_base_eval05-005.png"><img src="lora_model_ablation_report_images/flux2_klein_base_eval05-005.png" alt="Flux Klein output for eval05.jpg"></a>
             <figcaption><strong>Flux Klein</strong><span>768x768 · 0.39 MB</span></figcaption>
           </figure>
           <figure class="result-cell">
             <a href="lora_model_ablation_report_images/sd15_eval05-005.png"><img src="lora_model_ablation_report_images/sd15_eval05-005.png" alt="SD 1.5 output for eval05.jpg"></a>
             <figcaption><strong>SD 1.5</strong><span>512x512 · 0.26 MB</span></figcaption>
           </figure>
           <figure class="result-cell">
             <a href="lora_model_ablation_report_images/sd35_medium_eval05-005.png"><img src="lora_model_ablation_report_images/sd35_medium_eval05-005.png" alt="SD 3.5 output for eval05.jpg"></a>
             <figcaption><strong>SD 3.5</strong><span>1024x1024 · 0.97 MB</span></figcaption>
           </figure>
         </div>
       </section>
     </section>
   </main>
   <footer>
+    Generated from <code>ABLATION.md</code>, <code>dataset/train</code>, <code>dataset/conditioning</code>, <code>dataset/eval</code>, <code>outputs/ablation/local_edit_eval</code>, and <code>outputs/lora/base_muon/eval_outputs</code>. Image paths are repo-relative from this report location.
   </footer>
 </body>
 </html>