akhaliq HF Staff commited on
Commit
df3bdaf
·
verified ·
1 Parent(s): 3251e92

Add FAQ (need-a-LoRA, standalone adapter files, gaze, commercial terms)

Browse files
Files changed (1) hide show
  1. README.md +46 -0
README.md CHANGED
@@ -75,6 +75,49 @@ outfit detail.
75
 
76
  Full test set (8 hard cases: profile, small face, occlusion, full-body, group): [akhaliq/h3-image-swap-tests](https://huggingface.co/datasets/akhaliq/h3-image-swap-tests).
77
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
78
  ## ComfyUI
79
 
80
  One file, no LoRA nodes, no adapters — character swap is built into the checkpoint itself.
@@ -110,6 +153,7 @@ character-swap built in. Fusion provenance: [`comfyui/minimax_h3_image_character
110
 
111
  Because the adapter is fused in, there is nothing else to install: no LoRA node, no strength to
112
  tune — every image you edit with this checkpoint has character swap built in.
 
113
 
114
  *Note: end-to-end runs on this page were validated through the diffusers pipeline; the ComfyUI
115
  checkpoint is the identical math in Comfy's own layout. Report any ComfyUI-specific surprises in
@@ -177,6 +221,8 @@ on one H200. The merged DiT is a drop-in replacement anywhere
177
 
178
  - The LoRA was trained for **video** character swap (on Ref2VA, 24 fps clips); the port to images
179
  is validated on 8 hard cases, not a broad benchmark. Report surprises in the discussions tab.
 
 
180
  - Like the base model, it can drift on extreme poses, heavy occlusion, or very small faces.
181
  - The character-swap LoRA is an experimental community model; identity fidelity varies with the
182
  reference quality (a clean, multi-view character sheet works best).
 
75
 
76
  Full test set (8 hard cases: profile, small face, occlusion, full-body, group): [akhaliq/h3-image-swap-tests](https://huggingface.co/datasets/akhaliq/h3-image-swap-tests).
77
 
78
+ ## FAQ
79
+
80
+ **“Do you even need a LoRA for this? Doesn't H3-Image do swaps with the right prompt?”**
81
+ Often, yes — that was our first test too. Native H3-Image passes 7 of 8 hard cases (profile faces,
82
+ occlusion, small faces, full-body, group scenes) with a good prompt. It fails *systematically* on
83
+ cross-style swaps onto non-frontal faces: anime character → strict-profile photo failed on 3/3
84
+ seeds (a hybrid face, the character printed into a prop, and the swap ignored). That case — plus
85
+ tighter identity fidelity overall — is what this checkpoint adds. If your swaps are frontal and
86
+ photoreal, try the stock model with our prompt template first; if you hit its edge cases, this is
87
+ for you.
88
+
89
+ **“Where's the LoRA? Why is the ‘LoRA’ 40 GB?”**
90
+ It isn't a 40 GB adapter — this repo is the adapter *fused into* the full checkpoint (same size as
91
+ Comfy-Org's stock DiT), so there's one file and nothing to configure. If you want the 155 MB
92
+ adapter by itself:
93
+
94
+ | File | Layout | Use |
95
+ | --- | --- | --- |
96
+ | [`lora/h3_character_swap_v1.5.safetensors`](lora/h3_character_swap_v1.5.safetensors) (155 MB) | ComfyUI (akatz's original, unchanged) | `models/loras/` + **LoraLoaderModelOnly** @ strength 1.0 on the stock H3-Image DiT — all 208 targets verified to match |
97
+ | [`lora/h3_image_character_swap_diffusers.safetensors`](lora/h3_image_character_swap_diffusers.safetensors) (173 MB) | diffusers | `transformer.load_lora_adapter(state_dict, prefix=None)` for strength control |
98
+
99
+ No extraction needed — the merged checkpoint is exactly `W + B·A` of these adapters.
100
+
101
+ **“In some examples the character looks the wrong way.”**
102
+ That's the design, not a bug: a swap follows the *source* person's pose, position and gaze (it's
103
+ what makes it a swap instead of a paste). The model re-renders the reference character facing the
104
+ source's way; on extreme profile/cross-style combos that re-render can look off. Rephrase the
105
+ “same pose, position, scale and gaze” part of the prompt if you'd rather follow the reference's
106
+ orientation.
107
+
108
+ **“Can I use it commercially?”**
109
+ The checkpoint is under the [MiniMax-H3-Image Community License](LICENSE). Free, including
110
+ commercial use, in the Applicable Territory — which is worldwide **except the EU, UK, South Korea
111
+ and the USA**. Two triggers require MiniMax's prior written authorization (apply
112
+ [here](https://platform.minimax.io/console/license)): annualized revenue above **$1M**, or
113
+ providing it as a **hosted service** to third parties. In the excluded territories the community
114
+ license grants no rights — contact MiniMax. MiniMax claims no rights over your outputs. (This is a
115
+ summary, not legal advice — read the LICENSE.)
116
+
117
+ **“What prompt should I use?”**
118
+ See [Prompting tips](#comfyui) — scene as `<Picture 1>`, character as `<Picture 2>`, name the
119
+ person, and say what to preserve.
120
+
121
  ## ComfyUI
122
 
123
  One file, no LoRA nodes, no adapters — character swap is built into the checkpoint itself.
 
153
 
154
  Because the adapter is fused in, there is nothing else to install: no LoRA node, no strength to
155
  tune — every image you edit with this checkpoint has character swap built in.
156
+ (Prefer the 155 MB adapter by itself with strength control? See [the FAQ](#faq) — it's in [`lora/`](lora/).)
157
 
158
  *Note: end-to-end runs on this page were validated through the diffusers pipeline; the ComfyUI
159
  checkpoint is the identical math in Comfy's own layout. Report any ComfyUI-specific surprises in
 
221
 
222
  - The LoRA was trained for **video** character swap (on Ref2VA, 24 fps clips); the port to images
223
  is validated on 8 hard cases, not a broad benchmark. Report surprises in the discussions tab.
224
+ - Swaps follow the source person's pose and gaze by design; on extreme profile/cross-style combos
225
+ the re-rendered viewing direction can look off — tune the pose/gaze phrase in the prompt.
226
  - Like the base model, it can drift on extreme poses, heavy occlusion, or very small faces.
227
  - The character-swap LoRA is an experimental community model; identity fidelity varies with the
228
  reference quality (a clean, multi-view character sheet works best).