Mark the enlarge-then-fix hands way a fallback

In the user's tests Codex's FLUX-Fill redraw of one crop per hand did better.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

CHANGELOG.md

@@ -5,8 +5,8 @@
 - 跑图前的尺寸提醒放宽一档：6GB 显卡采样到约 420 万像素才提醒（以前 210 万），12GB 约 840 万。今天在 6GB 的笔记本显卡上实测，400 万像素一次约 100 秒，并不吃紧；以前的门槛会让 AI 动不动就问你要不要停。解码大图的提醒不变。
 - The size warning before a run is less eager: on a 6 GB card from about 4.2 megapixels (2.1 before), 8.4 on 12 GB. A 6 GB laptop card sampled 4 megapixels in about 100 s today; the old line had the AI ask you to stop runs that were fine. The warning for decoding a large picture is as before.
 
-- 新做法 **修手·先放大再修**：你递给 AI 一张手画坏的图，AMB 先把整张图放大到约 400 万像素（原图越小放得越大：1280×832 这种放 2 倍，已经很大的图不放，最多 2 倍）轻轻重画一遍，让手变大、模型有地方画手指，再自动找出每只手、放大重画后贴回去。这是照着参考作者的做法，在 6GB 的笔记本显卡上一组组试出来的：只修手不先放大，大约一半能成；先放大 1.5 倍再修，大约 4 张里 3 张能成；放大 2 倍，同样的种子手指更清楚。Anima 模型上有 Anima Turbo LoRA（做法面板里可以下载）时只跑 10 步，1280×832 的图约 100 秒；没有也能用，慢三倍左右。AI 会一次出 4 张，先自己放大看手，把手指数不对、粘在一起、大小不对的去掉，都不行就换种子再来，最多两轮，再让你在只放大手那一块的面板里挑。整张图会有些细小变化（脸、水花），构图不变。
-- New way **Fix hands: enlarge, then fix**: from a picture with bad hands, Anomalous enlarges the whole picture to about 4 megapixels (the smaller it is, the more: 2× for 1280×832, not at all for a large one, at most 2×) and redraws it lightly, so the hands grow and the model has room for fingers, then finds every hand and redraws it larger. Taken from the reference author's workflows and tried step by step on a 6 GB laptop card: the hand pass alone fixed about half the tries, enlarging 1.5× first about 3 in 4, and 2× gave sharper fingers in the same seeds. On an Anima model with the Anima Turbo LoRA (downloadable in the ways panel) it is 10 steps, about 100 s for 1280×832; without it, about three times slower. The AI makes 4 tries, checks the hands itself first, drops wrong finger counts, fused fingers and wrong sizes, tries new seeds at most twice more, and lets you pick in a panel zoomed on the hand. Small things change over the picture (faces, splashes); the layout stays.
+- 新做法 **修手·先放大再修（备选）**：你递给 AI 一张手画坏的图，AMB 先把整张图放大到约 400 万像素（原图越小放得越大：1280×832 这种放 2 倍，已经很大的图不放，最多 2 倍）轻轻重画一遍，让手变大、模型有地方画手指，再自动找出每只手、放大重画后贴回去。这是照着参考作者的做法，在 6GB 的笔记本显卡上一组组试出来的：只修手不先放大，大约一半能成；先放大 1.5 倍再修，大约 4 张里 3 张能成；放大 2 倍，同样的种子手指更清楚。Anima 模型上有 Anima Turbo LoRA（做法面板里可以下载）时只跑 10 步，1280×832 的图约 100 秒；没有也能用，慢三倍左右。AI 会一次出 4 张，先自己放大看手，把手指数不对、粘在一起、大小不对的去掉，都不行就换种子再来，最多两轮，再让你在只放大手那一块的面板里挑。整张图会有些细小变化（脸、水花），构图不变。实测不如 Codex 用 FLUX-Fill 每只手单独裁一块重画的办法，所以只作备选：AI 手里有那个办法时先用那个。
+- New way **Fix hands: enlarge, then fix (fallback)**: from a picture with bad hands, Anomalous enlarges the whole picture to about 4 megapixels (the smaller it is, the more: 2× for 1280×832, not at all for a large one, at most 2×) and redraws it lightly, so the hands grow and the model has room for fingers, then finds every hand and redraws it larger. Taken from the reference author's workflows and tried step by step on a 6 GB laptop card: the hand pass alone fixed about half the tries, enlarging 1.5× first about 3 in 4, and 2× gave sharper fingers in the same seeds. On an Anima model with the Anima Turbo LoRA (downloadable in the ways panel) it is 10 steps, about 100 s for 1280×832; without it, about three times slower. The AI makes 4 tries, checks the hands itself first, drops wrong finger counts, fused fingers and wrong sizes, tries new seeds at most twice more, and lets you pick in a panel zoomed on the hand. Small things change over the picture (faces, splashes); the layout stays. In the user's tests it did worse than Codex's FLUX-Fill redraw of one crop per hand, so it is a fallback: the AI uses that one when it has it.
 
 - **修一处（比如一只手）更容易成**：从 Codex 这几天修手的试验看，重画的力度太大时，新画的手会撑满你圈的范围，于是手越修越大。现在“局部重绘”默认力度降到修东西用的 0.45（官方节点那块 0.5），AI 也会被告知：圈得贴着手、保持低力度、一次出 4 张，再用一个**只放大手那一块**的挑选面板给你挑，挑完用同样放大的前后对比给你看。展台上的“挑一张”和前后对比滑块都能只显示每张图的同一小块。
 - **Fixing one spot (a hand) works more often**: Codex's hand tests showed that with a strong repaint the new hand fills the whole marked area and grows each time. The repaint block's strength is now 0.45 by default, a fix (0.5 for the built-in one), and the AI is told to mark tightly, keep it low, make 4 tries and let you pick in a picks panel **zoomed on the hand**, then show before and after zoomed the same. The stage's picks and slider can show the same small part of every picture.

api/blocks/fix_hands.json

@@ -2,15 +2,15 @@
   "job": "detail",
   "want": "hands",
   "rank": 2,
-  "short": "enlarge, then hands",
+  "short": "enlarge, then hands (fallback)",
   "enlarge": {
     "megapixels": 4.2,
     "most": 2.0,
     "unknown": 1.5
   },
   "title": {
-    "zh": "修手·先放大再修",
-    "en": "Fix hands: enlarge, then fix"
+    "zh": "修手·先放大再修（备选）",
+    "en": "Fix hands: enlarge, then fix (fallback)"
   },
   "about": {
     "zh": "整张图放大到约 400 万像素（小图 2 倍）轻轻重画，让手变大；再找出每只手，放大重画后贴回去。",
@@ -124,7 +124,7 @@
       }
     }
   },
-  "after": "The picture is enlarged to about 4 megapixels (2× for one of about 1 megapixel, at most 2×, not at all for a large one) and drawn again lightly (the sampler now takes it, at denoise 0.4), then every hand found is redrawn larger (hands, denoise 0.45) and saved with prefix hands_fixed; the enlarged one is saved too, without the hand pass. About 3 tries in 4 give a good hand: run_workflow with batch 4 (each try its own seed), then check them before the user does: zoom_image each hands_fixed picture on the hand, compare_with the old picture, and drop a hand with the wrong number of fingers, fused or melted fingers, or one far bigger or smaller than its arm. None good: run_workflow batch 4 again, at most twice. Then a picks panel zoomed on the hand lets the user choose among the good ones.",
+  "after": "A fallback: a FLUX-Fill redraw of one square crop around each hand did better in the user's tests; use that when you have it. The picture is enlarged to about 4 megapixels (2× for one of about 1 megapixel, at most 2×, not at all for a large one) and drawn again lightly (the sampler now takes it, at denoise 0.4), then every hand found is redrawn larger (hands, denoise 0.45) and saved with prefix hands_fixed; the enlarged one is saved too, without the hand pass. About 3 tries in 4 give a good hand: run_workflow with batch 4 (each try its own seed), then check them before the user does: zoom_image each hands_fixed picture on the hand, compare_with the old picture, and drop a hand with the wrong number of fingers, fused or melted fingers, or one far bigger or smaller than its arm. None good: run_workflow batch 4 again, at most twice. Then a picks panel zoomed on the hand lets the user choose among the good ones.",
   "notes": "Tried on the Codex test picture (1280×832, three characters, the hand about 110×78 px): the hand pass alone gave a good hand in about half the seeds whatever its size (512 or 768) and denoise (0.45 or 0.55); enlarging 1.5× first gave good hands in 3 of 4, 2× gave sharper fingers in the same seeds (about 105 s instead of 75 s). Denoise 0.4 for the enlarged redraw kept the layout; faces and splashes change a little.",
   "models": [
     "hand_yolov8s",
@@ -264,7 +264,7 @@
           }
         }
       },
-      "after": "On this Anima model the Anima Turbo LoRA (0.7) makes both passes 10 steps at CFG 1, about 105 s for 1280×832 (2×) on a 6 GB card. The picture is enlarged to about 4 megapixels (2× for one of about 1 megapixel, at most 2×, not at all for a large one) and drawn again lightly (the sampler now takes it, at denoise 0.4), then every hand found is redrawn larger (hands, denoise 0.45) and saved with prefix hands_fixed; the enlarged one is saved too, without the hand pass. About 3 tries in 4 give a good hand: run_workflow with batch 4 (each try its own seed), then check them before the user does: zoom_image each hands_fixed picture on the hand, compare_with the old picture, and drop a hand with the wrong number of fingers, fused or melted fingers, or one far bigger or smaller than its arm. None good: run_workflow batch 4 again, at most twice. Then a picks panel zoomed on the hand lets the user choose among the good ones."
+      "after": "A fallback: a FLUX-Fill redraw of one square crop around each hand did better in the user's tests; use that when you have it. On this Anima model the Anima Turbo LoRA (0.7) makes both passes 10 steps at CFG 1, about 105 s for 1280×832 (2×) on a 6 GB card. The picture is enlarged to about 4 megapixels (2× for one of about 1 megapixel, at most 2×, not at all for a large one) and drawn again lightly (the sampler now takes it, at denoise 0.4), then every hand found is redrawn larger (hands, denoise 0.45) and saved with prefix hands_fixed; the enlarged one is saved too, without the hand pass. About 3 tries in 4 give a good hand: run_workflow with batch 4 (each try its own seed), then check them before the user does: zoom_image each hands_fixed picture on the hand, compare_with the old picture, and drop a hand with the wrong number of fingers, fused or melted fingers, or one far bigger or smaller than its arm. None good: run_workflow batch 4 again, at most twice. Then a picks panel zoomed on the hand lets the user choose among the good ones."
     }
   }
 }