ComfyUI custom node (unmerged fp32 LoRA + fp32 time path + data-space sampler), example workflows, GPU check

.gitattributes

@@ -40,3 +40,4 @@ images/training/000300.jpg filter=lfs diff=lfs merge=lfs -text
 images/training/000500.jpg filter=lfs diff=lfs merge=lfs -text
 images/training/000800.jpg filter=lfs diff=lfs merge=lfs -text
 images/training/init.jpg filter=lfs diff=lfs merge=lfs -text
+images/comfyui-check.jpg filter=lfs diff=lfs merge=lfs -text

README.md

@@ -69,6 +69,26 @@ for r in range(4):                                        # 1〜3 回で止め
 - time path の 3 モジュールは fp32 のまま使います。
 - CFG とネガティブプロンプトは使いません。プロンプトの形式(自然文・タグ列・JSON レイアウト、日本語の文字指定)は MOSS-Image と同じです。
 
+### ComfyUI
+
+カスタムノード `comfyui/rapid-moss/` を同梱しています。ComfyUI 標準の `LoraLoaderModelOnly` は LoRA を重みにマージするので、このアダプタには使えません。このノードを使ってください。
+
+1. MOSS-Image 公式ノード([OpenMOSS/MOSS-Image](https://github.com/OpenMOSS/MOSS-Image) の `comfyui/moss-image/`)を入れます。
+2. このリポジトリの `comfyui/rapid-moss/` を `ComfyUI/custom_nodes/rapid-moss/` にコピーします。
+3. 上の 3 ファイルを `ComfyUI/models/rapid_moss/rapid-moss-4step/` に置きます。
+4. `comfyui/rapid-moss/example_workflows/rapid-moss-4step.json` を画面にドラッグ&ドロップし、HF Loader の `model_path` を MOSS-Image-SFT のフォルダにして実行します。
+
+ノード **MOSS-Image Few-Step Adapter (rapid-moss)** は、次の処理をして MODEL と、専用の SAMPLER・SIGMAS を返します。
+
+- LoRA をマージせず fp32 で足す
+- time path を fp32 に置き換える
+
+これを `BasicGuider`(CFG なし)と `SamplerCustomAdvanced` につなぎます。lodestones の 8 ステップ版も、同じ 3 ファイルを別フォルダに置けば同じノードで使えます。
+
+詳しい手順と注意点は [comfyui/rapid-moss/README.md](comfyui/rapid-moss/README.md) にあります。GPU での動作確認では、学習コードで作った画像とほぼ同じ絵になりました(下の図)。1024² 1 枚は 4.5〜6.9 秒(RTX PRO 6000、テキストエンコード・VAE 込み)です。
+
+![ComfyUI check](images/comfyui-check.jpg)
+
 ## 結果
 
 左から teacher(50 step)/ lodestones 8 step 版を 4 step で使ったもの / このアダプタ(4 step)。同じ seed、評価用 17 プロンプト(`eval/eval_prompts.jsonl`、学習には未使用)。

comfyui/rapid-moss/README.md

@@ -0,0 +1,85 @@
+# rapid-moss for ComfyUI
+
+MOSS-Image-SFT を **4 ステップ・CFG なし**で動かすカスタムノードです(rapid-moss の DMAD 4 ステップアダプタ。lodestones の 8 ステップ DMD も同じノードで使えます)。
+ComfyUI 標準の `SamplerCustomAdvanced` と組み合わせて使います。
+
+## 必要なもの
+
+| もの | 入手先 |
+|---|---|
+| ComfyUI | 新しめの版(`SamplerCustomAdvanced`、`BasicGuider`、`RandomNoise`、`EmptyFlux2LatentImage` があるもの。2026-10-16 の master で動作確認) |
+| MOSS-Image 公式ノード | [OpenMOSS/MOSS-Image](https://github.com/OpenMOSS/MOSS-Image) の `comfyui/moss-image/` を `ComfyUI/custom_nodes/moss-image/` へコピー(ComfyUI Manager で「MOSS-Image」を検索してもよい) |
+| MOSS-Image-SFT 本体 | [OpenMOSS-Team/MOSS-Image-SFT](https://huggingface.co/OpenMOSS-Team/MOSS-Image-SFT) のフォルダ一式(`transformer/`、`text_encoder/`、`tokenizer/`、`vae/`) |
+| このノード | このフォルダ(`comfyui/rapid-moss/`)を `ComfyUI/custom_nodes/rapid-moss/` へコピー |
+| アダプタ | [darask0/rapid-moss](https://huggingface.co/darask0/rapid-moss) の 3 ファイルを `ComfyUI/models/rapid_moss/rapid-moss-4step/` へ |
+
+```bash
+cd ComfyUI
+hf download OpenMOSS-Team/MOSS-Image-SFT --local-dir /path/to/MOSS-Image-SFT --exclude "assets/*"
+hf download darask0/rapid-moss lora.safetensors time_embedding.safetensors dataspace.safetensors \
+    --local-dir models/rapid_moss/rapid-moss-4step
+hf download darask0/rapid-moss --include "comfyui/rapid-moss/*" --local-dir /tmp/rapid-moss
+cp -r /tmp/rapid-moss/comfyui/rapid-moss custom_nodes/rapid-moss
+# 任意: lodestones の 8 ステップ版も使う場合
+hf download lodestones/MOSS-Image-SFT-8-steps-dmd lora.safetensors time_embedding.safetensors dataspace.safetensors \
+    --local-dir models/rapid_moss/lodestones-8step
+```
+
+`models/rapid_moss/` の下の、3 ファイル(`lora.safetensors`、`time_embedding.safetensors`、`dataspace.safetensors`)を含むフォルダが、ノードの `adapter` の選択肢に出ます。
+
+## 使い方
+
+1. ComfyUI を再起動し、`example_workflows/rapid-moss-4step.json` を画面にドラッグ&ドロップします。
+2. **MOSS-Image HF Loader (SFT)** の `model_path` に、MOSS-Image-SFT のフォルダを指定します。
+3. **Prompt** を書き換えて **実行** します。
+
+ノードのつなぎ方:
+
+```
+MOSS-Image HF Loader ─ MODEL ─▶ MOSS-Image Few-Step Adapter (rapid-moss) ─ model ──▶ BasicGuider ─▶ SamplerCustomAdvanced ─▶ VAE Decode ─▶ Save Image
+        │                                                        ├ sampler ───────────────────────────▲ (sampler)
+        ├ CLIP ─▶ CLIPTextEncode(プロンプト)─▶ BasicGuider        └ sigmas ────────────────────────────▲ (sigmas)
+        └ VAE ─────────────────────────────────────────────────────────────▶ VAE Decode
+RandomNoise ─▶ SamplerCustomAdvanced (noise)    EmptyFlux2LatentImage ─▶ SamplerCustomAdvanced (latent_image)
+```
+
+API で投げる場合は `example_workflows/rapid-moss-4step-api.json` を使います(`model_path` を書き換えてください)。
+
+## ノードの設定
+
+| 入力 | 説明 |
+|---|---|
+| `adapter` | `models/rapid_moss/` の下のフォルダ名 |
+| `steps` | ネットワークの計算回数。rapid-moss は 1〜4(4 推奨)、lodestones は 1〜8 |
+| `lora_strength` | 高速化 LoRA の強さ。1.0 推奨 |
+
+出力の `model`・`sampler`・`sigmas` は必ずセットで `BasicGuider` と `SamplerCustomAdvanced` に渡してください。
+
+## 注意
+
+- **CFG とネガティブプロンプトは使いません。** `BasicGuider` を使ってください。`CFGGuider` や `KSampler` で CFG を 1 より大きくすると、計算が倍になるうえ絵が崩れます。
+- **`ModelSamplingSD3`(shift)と `SkipLayerGuidanceDiTSimple`(layer 22 skip)は入れないでください。** 通常の 50 ステップ生成用の設定です。
+- **解像度:** `EmptyFlux2LatentImage` で 16 の倍数、約 100 万画素がおすすめです。学習した解像度は 1024×1024、896×1152、832×1216、768×1344 と、それぞれの横向きです。
+- **`weight_dtype`:** `default`(bf16)か `fp32` にしてください。fp8 には対応していません(エラーになります)。
+- **仕組み:** LoRA は重みにマージせず fp32 で足し、time path も fp32 に置き換えます。差分が bf16 の丸め幅より小さく、マージすると効果が消えるためです。このぶん VRAM を約 1.5 GB 余分に使います。
+- **seed:** `RandomNoise` の seed は、HF の `inference.py --seed` と同じ初期ノイズになります(どちらも CPU の `torch.randn`)。
+- **他の LoRA との併用:** 画風 LoRA などは `LoraLoaderModelOnly` で重ねられます。このノードの前でも後でも構いません。ただし 4 ステップ化した状態での効き方は検証していません。
+
+## 動作確認(2026-10-17)
+
+- **環境:** ComfyUI master(fc04da9、2026-10-16)、MOSS-Image 公式ノード(c1355e6)、torch 2.11 + CUDA 12.8、RTX PRO 6000。
+- **結果:** 学習コードで作った正解画像(同じ seed 1234)とほぼ同じ絵になりました。テキストエンコーダの実装の違いによる細かい差だけです。
+- **lodestones 8 ステップ:** 同じノードで動きました。
+- **時間:** 1024² 1 枚あたり 4.5〜6.9 秒(テキストエンコード・VAE 込み、他のジョブと GPU を共用した状態)。
+
+![ComfyUI check](https://huggingface.co/darask0/rapid-moss/resolve/main/images/comfyui-check.jpg)
+
+## English
+
+Custom node for 4-step, CFG-free MOSS-Image-SFT with the rapid-moss adapter (also runs lodestones' 8-step DMD). Install the
+official MOSS-Image nodes, copy this folder to `ComfyUI/custom_nodes/rapid-moss/`, and put `lora.safetensors`,
+`time_embedding.safetensors`, `dataspace.safetensors` in `ComfyUI/models/rapid_moss/<name>/`. Load
+`example_workflows/rapid-moss-4step.json`, set the HF loader's `model_path` to MOSS-Image-SFT, and run. Use `BasicGuider`
+(no CFG / negative), no `ModelSamplingSD3` or SLG, `weight_dtype` default/bf16/fp32 (fp8 unsupported). The node applies the
+LoRA unmerged in fp32 and swaps in the fp32 time path (~1.5 GB extra VRAM), and returns the matching SAMPLER and SIGMAS for
+`SamplerCustomAdvanced`.

comfyui/rapid-moss/__init__.py

@@ -0,0 +1,3 @@
+from .nodes import NODE_CLASS_MAPPINGS, NODE_DISPLAY_NAME_MAPPINGS
+
+__all__ = ['NODE_CLASS_MAPPINGS', 'NODE_DISPLAY_NAME_MAPPINGS']

comfyui/rapid-moss/example_workflows/rapid-moss-4step-api.json

@@ -0,0 +1,17 @@
+{
+  "1": {"class_type": "MOSSImageHFLoader", "_meta": {"title": "MOSS-Image HF Loader (SFT)"},
+        "inputs": {"model_path": "/path/to/MOSS-Image-SFT", "weight_dtype": "default", "text_device": "default"}},
+  "2": {"class_type": "RapidMOSSAdapter", "_meta": {"title": "MOSS-Image Few-Step Adapter (rapid-moss)"},
+        "inputs": {"model": ["1", 0], "adapter": "rapid-moss-4step", "steps": 4, "lora_strength": 1.0}},
+  "3": {"class_type": "CLIPTextEncode", "_meta": {"title": "Prompt"},
+        "inputs": {"clip": ["1", 1], "text": "A high-quality anime illustration of a girl with long silver hair and teal eyes, wearing a yellow raincoat, standing at a rainy bus stop at night, neon reflections on wet asphalt"}},
+  "4": {"class_type": "BasicGuider", "_meta": {"title": "BasicGuider (no CFG)"},
+        "inputs": {"model": ["2", 0], "conditioning": ["3", 0]}},
+  "5": {"class_type": "RandomNoise", "_meta": {"title": "RandomNoise"}, "inputs": {"noise_seed": 42}},
+  "6": {"class_type": "EmptyFlux2LatentImage", "_meta": {"title": "Empty Flux 2 Latent"},
+        "inputs": {"width": 1024, "height": 1024, "batch_size": 1}},
+  "7": {"class_type": "SamplerCustomAdvanced", "_meta": {"title": "SamplerCustomAdvanced"},
+        "inputs": {"noise": ["5", 0], "guider": ["4", 0], "sampler": ["2", 1], "sigmas": ["2", 2], "latent_image": ["6", 0]}},
+  "8": {"class_type": "VAEDecode", "_meta": {"title": "VAE Decode"}, "inputs": {"samples": ["7", 0], "vae": ["1", 2]}},
+  "9": {"class_type": "SaveImage", "_meta": {"title": "Save Image"}, "inputs": {"filename_prefix": "rapid-moss", "images": ["8", 0]}}
+}

comfyui/rapid-moss/example_workflows/rapid-moss-4step.json

@@ -0,0 +1,495 @@
+{
+ "last_node_id": 9,
+ "last_link_id": 12,
+ "nodes": [
+  {
+   "id": 1,
+   "type": "MOSSImageHFLoader",
+   "pos": [
+    40,
+    120
+   ],
+   "size": [
+    380,
+    130
+   ],
+   "flags": {},
+   "order": 0,
+   "mode": 0,
+   "inputs": [],
+   "outputs": [
+    {
+     "name": "MODEL",
+     "type": "MODEL",
+     "links": [
+      1
+     ],
+     "slot_index": 0
+    },
+    {
+     "name": "CLIP",
+     "type": "CLIP",
+     "links": [
+      2
+     ],
+     "slot_index": 1
+    },
+    {
+     "name": "VAE",
+     "type": "VAE",
+     "links": [
+      11
+     ],
+     "slot_index": 2
+    }
+   ],
+   "properties": {
+    "Node name for S&R": "MOSSImageHFLoader"
+   },
+   "widgets_values": [
+    "/path/to/MOSS-Image-SFT",
+    "default",
+    "default"
+   ],
+   "title": "MOSS-Image HF Loader (SFT)"
+  },
+  {
+   "id": 2,
+   "type": "RapidMOSSAdapter",
+   "pos": [
+    470,
+    60
+   ],
+   "size": [
+    380,
+    130
+   ],
+   "flags": {},
+   "order": 1,
+   "mode": 0,
+   "inputs": [
+    {
+     "name": "model",
+     "type": "MODEL",
+     "link": 1
+    }
+   ],
+   "outputs": [
+    {
+     "name": "model",
+     "type": "MODEL",
+     "links": [
+      3
+     ],
+     "slot_index": 0
+    },
+    {
+     "name": "sampler",
+     "type": "SAMPLER",
+     "links": [
+      7
+     ],
+     "slot_index": 1
+    },
+    {
+     "name": "sigmas",
+     "type": "SIGMAS",
+     "links": [
+      8
+     ],
+     "slot_index": 2
+    }
+   ],
+   "properties": {
+    "Node name for S&R": "RapidMOSSAdapter"
+   },
+   "widgets_values": [
+    "rapid-moss-4step",
+    4,
+    1.0
+   ]
+  },
+  {
+   "id": 3,
+   "type": "CLIPTextEncode",
+   "pos": [
+    470,
+    250
+   ],
+   "size": [
+    420,
+    200
+   ],
+   "flags": {},
+   "order": 2,
+   "mode": 0,
+   "inputs": [
+    {
+     "name": "clip",
+     "type": "CLIP",
+     "link": 2
+    }
+   ],
+   "outputs": [
+    {
+     "name": "CONDITIONING",
+     "type": "CONDITIONING",
+     "links": [
+      4
+     ],
+     "slot_index": 0
+    }
+   ],
+   "properties": {
+    "Node name for S&R": "CLIPTextEncode"
+   },
+   "widgets_values": [
+    "A high-quality anime illustration of a girl with long silver hair and teal eyes, wearing a yellow raincoat, standing at a rainy bus stop at night, neon reflections on wet asphalt"
+   ],
+   "title": "Prompt"
+  },
+  {
+   "id": 4,
+   "type": "BasicGuider",
+   "pos": [
+    940,
+    60
+   ],
+   "size": [
+    260,
+    50
+   ],
+   "flags": {},
+   "order": 3,
+   "mode": 0,
+   "inputs": [
+    {
+     "name": "model",
+     "type": "MODEL",
+     "link": 3
+    },
+    {
+     "name": "conditioning",
+     "type": "CONDITIONING",
+     "link": 4
+    }
+   ],
+   "outputs": [
+    {
+     "name": "GUIDER",
+     "type": "GUIDER",
+     "links": [
+      6
+     ],
+     "slot_index": 0
+    }
+   ],
+   "properties": {
+    "Node name for S&R": "BasicGuider"
+   },
+   "widgets_values": [],
+   "title": "BasicGuider (no CFG)"
+  },
+  {
+   "id": 5,
+   "type": "RandomNoise",
+   "pos": [
+    940,
+    160
+   ],
+   "size": [
+    260,
+    82
+   ],
+   "flags": {},
+   "order": 4,
+   "mode": 0,
+   "inputs": [],
+   "outputs": [
+    {
+     "name": "NOISE",
+     "type": "NOISE",
+     "links": [
+      5
+     ],
+     "slot_index": 0
+    }
+   ],
+   "properties": {
+    "Node name for S&R": "RandomNoise"
+   },
+   "widgets_values": [
+    42,
+    "randomize"
+   ]
+  },
+  {
+   "id": 6,
+   "type": "EmptyFlux2LatentImage",
+   "pos": [
+    940,
+    290
+   ],
+   "size": [
+    260,
+    106
+   ],
+   "flags": {},
+   "order": 5,
+   "mode": 0,
+   "inputs": [],
+   "outputs": [
+    {
+     "name": "LATENT",
+     "type": "LATENT",
+     "links": [
+      9
+     ],
+     "slot_index": 0
+    }
+   ],
+   "properties": {
+    "Node name for S&R": "EmptyFlux2LatentImage"
+   },
+   "widgets_values": [
+    1024,
+    1024,
+    1
+   ]
+  },
+  {
+   "id": 7,
+   "type": "SamplerCustomAdvanced",
+   "pos": [
+    1250,
+    60
+   ],
+   "size": [
+    300,
+    110
+   ],
+   "flags": {},
+   "order": 6,
+   "mode": 0,
+   "inputs": [
+    {
+     "name": "noise",
+     "type": "NOISE",
+     "link": 5
+    },
+    {
+     "name": "guider",
+     "type": "GUIDER",
+     "link": 6
+    },
+    {
+     "name": "sampler",
+     "type": "SAMPLER",
+     "link": 7
+    },
+    {
+     "name": "sigmas",
+     "type": "SIGMAS",
+     "link": 8
+    },
+    {
+     "name": "latent_image",
+     "type": "LATENT",
+     "link": 9
+    }
+   ],
+   "outputs": [
+    {
+     "name": "output",
+     "type": "LATENT",
+     "links": [
+      10
+     ],
+     "slot_index": 0
+    },
+    {
+     "name": "denoised_output",
+     "type": "LATENT",
+     "links": [],
+     "slot_index": 1
+    }
+   ],
+   "properties": {
+    "Node name for S&R": "SamplerCustomAdvanced"
+   },
+   "widgets_values": []
+  },
+  {
+   "id": 8,
+   "type": "VAEDecode",
+   "pos": [
+    1250,
+    230
+   ],
+   "size": [
+    210,
+    46
+   ],
+   "flags": {},
+   "order": 7,
+   "mode": 0,
+   "inputs": [
+    {
+     "name": "samples",
+     "type": "LATENT",
+     "link": 10
+    },
+    {
+     "name": "vae",
+     "type": "VAE",
+     "link": 11
+    }
+   ],
+   "outputs": [
+    {
+     "name": "IMAGE",
+     "type": "IMAGE",
+     "links": [
+      12
+     ],
+     "slot_index": 0
+    }
+   ],
+   "properties": {
+    "Node name for S&R": "VAEDecode"
+   },
+   "widgets_values": []
+  },
+  {
+   "id": 9,
+   "type": "SaveImage",
+   "pos": [
+    1600,
+    60
+   ],
+   "size": [
+    520,
+    560
+   ],
+   "flags": {},
+   "order": 8,
+   "mode": 0,
+   "inputs": [
+    {
+     "name": "images",
+     "type": "IMAGE",
+     "link": 12
+    }
+   ],
+   "outputs": [],
+   "properties": {
+    "Node name for S&R": "SaveImage"
+   },
+   "widgets_values": [
+    "rapid-moss"
+   ]
+  }
+ ],
+ "links": [
+  [
+   1,
+   1,
+   0,
+   2,
+   0,
+   "MODEL"
+  ],
+  [
+   2,
+   1,
+   1,
+   3,
+   0,
+   "CLIP"
+  ],
+  [
+   3,
+   2,
+   0,
+   4,
+   0,
+   "MODEL"
+  ],
+  [
+   4,
+   3,
+   0,
+   4,
+   1,
+   "CONDITIONING"
+  ],
+  [
+   5,
+   5,
+   0,
+   7,
+   0,
+   "NOISE"
+  ],
+  [
+   6,
+   4,
+   0,
+   7,
+   1,
+   "GUIDER"
+  ],
+  [
+   7,
+   2,
+   1,
+   7,
+   2,
+   "SAMPLER"
+  ],
+  [
+   8,
+   2,
+   2,
+   7,
+   3,
+   "SIGMAS"
+  ],
+  [
+   9,
+   6,
+   0,
+   7,
+   4,
+   "LATENT"
+  ],
+  [
+   10,
+   7,
+   0,
+   8,
+   0,
+   "LATENT"
+  ],
+  [
+   11,
+   1,
+   2,
+   8,
+   1,
+   "VAE"
+  ],
+  [
+   12,
+   8,
+   0,
+   9,
+   0,
+   "IMAGE"
+  ]
+ ],
+ "groups": [],
+ "config": {},
+ "extra": {},
+ "version": 0.4
+}

comfyui/rapid-moss/nodes.py

@@ -0,0 +1,149 @@
+"""Few-step MOSS-Image adapters (rapid-moss DMAD 4-step, lodestones 8-step DMD) for ComfyUI.
+
+Works on the MODEL from the MOSS-Image nodes (github.com/OpenMOSS/MOSS-Image, comfyui/moss-image).
+An adapter is a folder in models/rapid_moss/ holding lora.safetensors, time_embedding.safetensors and
+dataspace.safetensors. The node returns a patched MODEL plus the data-space SAMPLER and SIGMAS for
+SamplerCustomAdvanced (use BasicGuider: no CFG, no negative prompt).
+"""
+import os
+
+import comfy.samplers
+import folder_paths
+import torch
+from safetensors.torch import load_file
+from torch.nn import functional as F
+
+FOLDER = 'rapid_moss'
+folder_paths.add_model_folder_path(FOLDER, os.path.join(folder_paths.models_dir, FOLDER))
+FILES = ('lora.safetensors', 'time_embedding.safetensors', 'dataspace.safetensors')
+PROJ = ('q', 'k', 'v', 'o', 'ffn_up_gate', 'ffn_down')
+DM = 'diffusion_model.'
+
+
+def _adapter_dirs():
+    found = {}
+    for root in folder_paths.get_folder_paths(FOLDER):
+        if os.path.isdir(root):
+            for name in sorted(os.listdir(root)):
+                path = os.path.join(root, name)
+                if name not in found and all(os.path.isfile(os.path.join(path, f)) for f in FILES):
+                    found[name] = path
+    return found
+
+
+class _DeviceCopy:
+    """fp32 tensors kept on the CPU, copied once to the device that first needs them."""
+
+    def __init__(self, *tensors):
+        self.cpu = tensors
+        self.device = None
+        self.copies = None
+
+    def get(self, device):
+        if device != self.device:
+            self.copies = [t.to(device) for t in self.cpu]
+            self.device = device
+        return self.copies
+
+
+class _FP32Linear(torch.nn.Module):
+    """fp32 linear outside ComfyUI's weight management. `weight` is a plain attribute so the MOSS
+    forward (which casts its input to `.weight.dtype`) runs this path in fp32."""
+
+    def __init__(self, weight, bias):
+        super().__init__()
+        self.weight = weight
+        self.tensors = _DeviceCopy(weight, bias)
+
+    def forward(self, x):
+        weight, bias = self.tensors.get(x.device)
+        return F.linear(x.float(), weight, bias)
+
+
+def _lora_forward(module, down, up, scale):
+    """Unmerged fp32 LoRA: the delta is below bf16 rounding, so it must not be merged into the weight."""
+    class_forward = type(module).forward
+    tensors = _DeviceCopy(down, up)
+
+    def forward(x, *args, **kwargs):
+        out = class_forward(module, x, *args, **kwargs)
+        d, u = tensors.get(x.device)
+        return out + (((x.float() @ d.t()) @ u.t()) * scale).to(out.dtype)
+    return forward
+
+
+@torch.no_grad()
+def sample_dataspace(model, x, sigmas, extra_args=None, callback=None, disable=None, Q=None, codes=None):
+    """x0 = 0; for r: x = z + Q x0, v = net(x, code_r), x0 = z - v (one noise z, no schedule, no CFG)."""
+    extra_args = {} if extra_args is None else extra_args
+    z = x.float()  # sigmas[0] == 1, so the sampler receives pure noise
+    mix = Q.to(z.device, torch.float32)
+    x0 = torch.zeros_like(z)
+    s_in = z.new_ones([z.shape[0]])
+    for r in range(len(sigmas) - 1):
+        x_in = z + torch.einsum('oc,bchw->bohw', mix, x0)
+        s = float(codes[r]) / 1000.0  # flow model: timestep = sigma * 1000 = code
+        denoised = model(x_in, s * s_in, **extra_args).float()  # = x_in - s * v
+        x0 = z - (x_in - denoised) / s
+        if callback is not None:
+            callback({'x': x_in, 'i': r, 'sigma': sigmas[r], 'sigma_hat': sigmas[r], 'denoised': x0})
+    return x0
+
+
+class RapidMOSSAdapter:
+    @classmethod
+    def INPUT_TYPES(cls):
+        names = list(_adapter_dirs()) or ['(no adapter folder in models/rapid_moss)']
+        return {'required': {
+            'model': ('MODEL',),
+            'adapter': (names,),
+            'steps': ('INT', {'default': 4, 'min': 1, 'max': 8,
+                              'tooltip': 'Network evaluations; up to the adapter step-code count (4 for rapid-moss, 8 for lodestones).'}),
+            'lora_strength': ('FLOAT', {'default': 1.0, 'min': 0.0, 'max': 2.0, 'step': 0.05}),
+        }}
+
+    RETURN_TYPES = ('MODEL', 'SAMPLER', 'SIGMAS')
+    RETURN_NAMES = ('model', 'sampler', 'sigmas')
+    FUNCTION = 'apply'
+    CATEGORY = 'MOSS-Image/rapid-moss'
+    DESCRIPTION = ('Few-step MOSS-Image: patches the MOSS-Image MODEL with a data-space adapter (unmerged fp32 LoRA, '
+                   'fp32 time path) and returns the matching SAMPLER and SIGMAS for SamplerCustomAdvanced + BasicGuider.')
+
+    def apply(self, model, adapter, steps, lora_strength):
+        dirs = _adapter_dirs()
+        if adapter not in dirs:
+            raise FileNotFoundError(f'Put {", ".join(FILES)} in a folder under models/{FOLDER}/ (got {adapter!r}).')
+        path = dirs[adapter]
+        lora = load_file(os.path.join(path, 'lora.safetensors'))
+        time_path = load_file(os.path.join(path, 'time_embedding.safetensors'))
+        data = load_file(os.path.join(path, 'dataspace.safetensors'))
+        codes = data['step_codes'].float()
+        if steps > len(codes):
+            raise ValueError(f'{adapter} has {len(codes)} step codes; steps must be <= {len(codes)}.')
+
+        patched = model.clone()
+        base = patched.get_model_object(DM + 'shared_block_modulation')
+        if base.weight.dtype not in (torch.bfloat16, torch.float16, torch.float32):
+            raise ValueError('Load MOSS-Image with weight_dtype default/bf16/fp32; fp8 weights are not supported here.')
+        patched.add_object_patch(DM + 'time_embedding', torch.nn.Sequential(
+            _FP32Linear(time_path['time_embedding.0.weight'].float(), time_path['time_embedding.0.bias'].float()),
+            torch.nn.SiLU(),
+            _FP32Linear(time_path['time_embedding.2.weight'].float(), time_path['time_embedding.2.bias'].float())))
+        for name in ('shared_block_modulation', 'video_head_modulation'):
+            weight = patched.get_model_object(DM + name).weight.detach().to('cpu', torch.float32)
+            patched.add_object_patch(DM + name, _FP32Linear(weight, time_path[name + '.bias'].float()))
+        for i in range(len(patched.get_model_object(DM + 'layers'))):
+            for proj in PROJ:
+                target = f'{DM}layers.{i}.proj.{proj}'
+                down = lora[target + '.lora_down.weight'].float()
+                up = lora[target + '.lora_up.weight'].float()
+                scale = lora_strength * float(lora[target + '.alpha']) / down.shape[0]
+                patched.add_object_patch(target + '.forward', _lora_forward(patched.get_model_object(target), down, up, scale))
+
+        sampler = comfy.samplers.KSAMPLER(sample_dataspace, extra_options={'Q': data['latent_mix.Q'].float(), 'codes': codes})
+        sigmas = torch.cat((codes[:steps] / 1000.0, torch.zeros(1)))
+        return (patched, sampler, sigmas)
+
+
+NODE_CLASS_MAPPINGS = {'RapidMOSSAdapter': RapidMOSSAdapter}
+NODE_DISPLAY_NAME_MAPPINGS = {'RapidMOSSAdapter': 'MOSS-Image Few-Step Adapter (rapid-moss)'}