RioShiina commited on
Commit
598748a
·
verified ·
1 Parent(s): d9dfe83

feat: add P-H3-Image architecture

Browse files
README.md CHANGED
@@ -127,6 +127,7 @@ models:
127
  - OnomaAIResearch/Illustrious-XL-v2.0
128
  - ostris/krea2_turbo_style_reference
129
  - Patil/Krea-2-depth-controlnet
 
130
  - RedRayz/hikari_noob_v-pred_1.2.4
131
  - Shakker-Labs/FLUX.1-dev-ControlNet-Union-Pro-2.0
132
  - silveroxides/Chroma1-Radiance-fp8-scaled
@@ -195,6 +196,7 @@ models:
195
  - OmniGen2/OmniGen2
196
  - OpenMOSS-Team/MOSS-Image-Base
197
  - OpenMOSS-Team/MOSS-Image-SFT
 
198
  - Qwen/Qwen3.8-27B
199
  - Qwen/Qwen-Image
200
  - Qwen/Qwen-Image-2512
 
127
  - OnomaAIResearch/Illustrious-XL-v2.0
128
  - ostris/krea2_turbo_style_reference
129
  - Patil/Krea-2-depth-controlnet
130
+ - PrunaAI/P-H3-Image-Comfy
131
  - RedRayz/hikari_noob_v-pred_1.2.4
132
  - Shakker-Labs/FLUX.1-dev-ControlNet-Union-Pro-2.0
133
  - silveroxides/Chroma1-Radiance-fp8-scaled
 
196
  - OmniGen2/OmniGen2
197
  - OpenMOSS-Team/MOSS-Image-Base
198
  - OpenMOSS-Team/MOSS-Image-SFT
199
+ - PrunaAI/P-H3-Image
200
  - Qwen/Qwen3.8-27B
201
  - Qwen/Qwen-Image
202
  - Qwen/Qwen-Image-2512
core/pipelines/workflow_recipes/_partials/conditioning/p-h3-image.yaml ADDED
@@ -0,0 +1,117 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ nodes:
2
+ unet_loader:
3
+ class_type: UNETLoader
4
+ title: "Load Diffusion Model"
5
+ params:
6
+ weight_dtype: "default"
7
+ vae_loader:
8
+ class_type: VAELoader
9
+ title: "Load VAE"
10
+ clip_loader:
11
+ class_type: CLIPLoader
12
+ title: "Load CLIP"
13
+ params:
14
+ type: "minimax"
15
+ device: "default"
16
+ sigma_shift:
17
+ class_type: MiniMaxH3SigmaShift
18
+ title: "ModelSamplingMiniMaxH3 (keep at 1.0)"
19
+ params:
20
+ shift_video: 1.0
21
+ shift_audio: 1.0
22
+ prompt:
23
+ class_type: PH3ImageConditioning
24
+ title: "p-h3-image Conditioning"
25
+ params:
26
+ width: 2912
27
+ height: 2912
28
+ guider:
29
+ class_type: BasicGuider
30
+ title: "Basic Guider"
31
+ sampler_select:
32
+ class_type: KSamplerSelect
33
+ title: "KSamplerSelect"
34
+ params:
35
+ sampler_name: "euler"
36
+ manual_sigmas:
37
+ class_type: ManualSigmas
38
+ title: "ManualSigmas (fixed 8-step grid)"
39
+ params:
40
+ sigmas: "1.0, 0.984753, 0.967905, 0.951688, 0.925372, 0.881923, 0.810083, 0.528718, 0.0"
41
+ random_noise:
42
+ class_type: RandomNoise
43
+ title: "RandomNoise"
44
+ # Overrides the stock KSampler from _base_sampler.yaml: P-H3-Image samples on a
45
+ # fixed ManualSigmas grid with no CFG, so SamplerCustomAdvanced is used instead.
46
+ ksampler:
47
+ class_type: SamplerCustomAdvanced
48
+ title: "SamplerCustomAdvanced"
49
+
50
+ connections:
51
+ - from: "unet_loader:0"
52
+ to: "sigma_shift:model"
53
+ - from: "sigma_shift:0"
54
+ to: "guider:model"
55
+
56
+ - from: "clip_loader:0"
57
+ to: "prompt:clip"
58
+ - from: "vae_loader:0"
59
+ to: "prompt:vae"
60
+
61
+ - from: "prompt:0"
62
+ to: "guider:conditioning"
63
+
64
+ - from: "random_noise:0"
65
+ to: "ksampler:noise"
66
+ - from: "guider:0"
67
+ to: "ksampler:guider"
68
+ - from: "sampler_select:0"
69
+ to: "ksampler:sampler"
70
+ - from: "manual_sigmas:0"
71
+ to: "ksampler:sigmas"
72
+ - from: "prompt:1"
73
+ to: "ksampler:latent_image"
74
+
75
+ - from: "ksampler:0"
76
+ to: "vae_decode:samples"
77
+ - from: "vae_loader:0"
78
+ to: "vae_decode:vae"
79
+ - from: "vae_loader:0"
80
+ to: "vae_encode:vae"
81
+
82
+ dynamic_lora_chains:
83
+ lora_chain:
84
+ template: "LoraLoader"
85
+ output_map:
86
+ "unet_loader:0": "model"
87
+ "clip_loader:0": "clip"
88
+ input_map:
89
+ "model": "model"
90
+ "clip": "clip"
91
+ end_input_map:
92
+ "model": ["sigma_shift:model"]
93
+ "clip": ["prompt:clip"]
94
+
95
+ dynamic_reference_image_chains:
96
+ reference_image_chain:
97
+ text_encode_node: "prompt"
98
+ vae_node: "vae_loader"
99
+ max_images: 8
100
+
101
+ # Steps / CFG are intentionally discarded: P-H3-Image is trained and sampled on a
102
+ # fixed 8-step ManualSigmas grid with cfg = 1 (no CFG / no negative prompt).
103
+ # Pointing these keys at the non-existent "dummy" node makes the assembler skip them.
104
+ ui_map:
105
+ positive_prompt: "prompt:prompt"
106
+ negative_prompt: "dummy:negative_prompt"
107
+ unet_name: "unet_loader:unet_name"
108
+ clip_name: "clip_loader:clip_name"
109
+ vae_name: "vae_loader:vae_name"
110
+ width: "prompt:width"
111
+ height: "prompt:height"
112
+ seed: "random_noise:noise_seed"
113
+ steps: "dummy:steps"
114
+ cfg: "dummy:cfg"
115
+ sampler_name: "dummy:sampler_name"
116
+ scheduler: "dummy:scheduler"
117
+ denoise: "dummy:denoise"
yaml/chain_features.yaml CHANGED
@@ -266,6 +266,21 @@ minimax_h3_image_reference_edit:
266
  required:
267
  - image
268
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
269
  krea2_identity_edit:
270
  chains: krea2_identity_edit
271
  display_name: "KREA2 Identity Edit"
 
266
  required:
267
  - image
268
 
269
+ p_h3_image_reference_edit:
270
+ chains: reference_image
271
+ display_name: "P-H3-Image Reference Edit"
272
+ description: "For P-H3-Image, this feature enables reference image editing and combining capabilities. In txt2img mode, adding a single reference image performs an Image Edit, while adding multiple images performs an Image Combine. The prompt refers to them as <Picture 1>, <Picture 2>, ..."
273
+ max_count: 8
274
+ usage_guideline: "Supply reference image(s) encoded as Base64 Data URI or HTTP/HTTPS URL (up to 8). Refer to them in the prompt as <Picture 1>, <Picture 2>, ...; <Picture 1> is the picture to edit. Note: steps and cfg are fixed by the model (8-step ManualSigmas grid, no CFG)."
275
+ parameters_schema:
276
+ type: object
277
+ properties:
278
+ image:
279
+ type: string
280
+ description: "Reference image encoded as Base64 Data URI or HTTP/HTTPS URL."
281
+ required:
282
+ - image
283
+
284
  krea2_identity_edit:
285
  chains: krea2_identity_edit
286
  display_name: "KREA2 Identity Edit"
yaml/constants.yaml CHANGED
@@ -32,6 +32,14 @@ RESOLUTION_MAP:
32
  "3:4 (Classic Portrait)": [1344, 1792]
33
  "3:2 (Photography)": [1920, 1280]
34
  "2:3 (Photography Portrait)": [1280, 1920]
 
 
 
 
 
 
 
 
35
  krea-2:
36
  "1:1 (Square)": [1024, 1024]
37
  "16:9 (Landscape)": [1344, 768]
 
32
  "3:4 (Classic Portrait)": [1344, 1792]
33
  "3:2 (Photography)": [1920, 1280]
34
  "2:3 (Photography Portrait)": [1280, 1920]
35
+ p-h3-image:
36
+ "1:1 (Square)": [2912, 2912]
37
+ "16:9 (Landscape)": [3872, 2176]
38
+ "9:16 (Portrait)": [2176, 3872]
39
+ "4:3 (Classic)": [3360, 2496]
40
+ "3:4 (Classic Portrait)": [2496, 3360]
41
+ "3:2 (Photography)": [3552, 2368]
42
+ "2:3 (Photography Portrait)": [2368, 3552]
43
  krea-2:
44
  "1:1 (Square)": [1024, 1024]
45
  "16:9 (Landscape)": [1344, 768]
yaml/file_list.yaml CHANGED
@@ -405,6 +405,11 @@ file:
405
  repo_id: "comfyanonymous/ControlNet-v1-1_fp16_safetensors"
406
  repository_file_path: "control_v11u_sd15_tile_fp16.safetensors"
407
  diffusion_models:
 
 
 
 
 
408
  # MiniMax-H3-Image
409
  - filename: "minimax_h3_image_pruned_w4a8.safetensors"
410
  source: "hf"
@@ -975,6 +980,11 @@ file:
975
  repo_id: "Comfy-Org/Qwen-Image-2.1"
976
  repository_file_path: "model_patches/qwen_image_2.1_fun_controlnet_union_int8_convrot.safetensors"
977
  text_encoders:
 
 
 
 
 
978
  # MiniMax-H3-Image
979
  - filename: "qwen3vl_32b_minimax_h3_nvfp4_awq.safetensors"
980
  source: "hf"
@@ -1129,6 +1139,11 @@ file:
1129
  repo_id: "Comfy-Org/Omnigen2_ComfyUI_repackaged"
1130
  repository_file_path: "split_files/text_encoders/qwen_2.5_vl_fp16.safetensors"
1131
  vae:
 
 
 
 
 
1132
  # MiniMax-H3-Image
1133
  - filename: "minimax_h3_image_vae_fp16.safetensors"
1134
  source: "hf"
 
405
  repo_id: "comfyanonymous/ControlNet-v1-1_fp16_safetensors"
406
  repository_file_path: "control_v11u_sd15_tile_fp16.safetensors"
407
  diffusion_models:
408
+ # P-H3-Image
409
+ - filename: "p-h3-image_w4a8.safetensors"
410
+ source: "hf"
411
+ repo_id: "PrunaAI/P-H3-Image-Comfy"
412
+ repository_file_path: "diffusion_models/p-h3-image_w4a8.safetensors"
413
  # MiniMax-H3-Image
414
  - filename: "minimax_h3_image_pruned_w4a8.safetensors"
415
  source: "hf"
 
980
  repo_id: "Comfy-Org/Qwen-Image-2.1"
981
  repository_file_path: "model_patches/qwen_image_2.1_fun_controlnet_union_int8_convrot.safetensors"
982
  text_encoders:
983
+ # P-H3-Image
984
+ - filename: "p-h3-image_text_encoder_w4a8.safetensors"
985
+ source: "hf"
986
+ repo_id: "PrunaAI/P-H3-Image-Comfy"
987
+ repository_file_path: "text_encoders/p-h3-image_text_encoder_w4a8.safetensors"
988
  # MiniMax-H3-Image
989
  - filename: "qwen3vl_32b_minimax_h3_nvfp4_awq.safetensors"
990
  source: "hf"
 
1139
  repo_id: "Comfy-Org/Omnigen2_ComfyUI_repackaged"
1140
  repository_file_path: "split_files/text_encoders/qwen_2.5_vl_fp16.safetensors"
1141
  vae:
1142
+ # P-H3-Image
1143
+ - filename: "p-h3-image_vae_int8_convrot.safetensors"
1144
+ source: "hf"
1145
+ repo_id: "PrunaAI/P-H3-Image-Comfy"
1146
+ repository_file_path: "vae/p-h3-image_vae_int8_convrot.safetensors"
1147
  # MiniMax-H3-Image
1148
  - filename: "minimax_h3_image_vae_fp16.safetensors"
1149
  source: "hf"
yaml/model_architecture_features.yaml CHANGED
@@ -17,6 +17,11 @@ minimax-h3-image:
17
  - lora
18
  - reference_image
19
 
 
 
 
 
 
20
  krea-2:
21
  enabled_chains:
22
  - lora
 
17
  - lora
18
  - reference_image
19
 
20
+ p-h3-image:
21
+ enabled_chains:
22
+ - lora
23
+ - reference_image
24
+
25
  krea-2:
26
  enabled_chains:
27
  - lora
yaml/model_architectures.yaml CHANGED
@@ -1,6 +1,7 @@
1
  architecture_order:
2
  - "FLUX.3"
3
  - "MiniMax-H3-Image"
 
4
  - "Krea-2"
5
  - "Qwen-Image-2.1"
6
  - "Fibo-2.0"
@@ -46,6 +47,8 @@ architectures:
46
  model_type: "flux3"
47
  "MiniMax-H3-Image":
48
  model_type: "minimax-h3-image"
 
 
49
  Krea-2:
50
  model_type: "krea-2"
51
  "SenseNova-U1.5":
 
1
  architecture_order:
2
  - "FLUX.3"
3
  - "MiniMax-H3-Image"
4
+ - "P-H3-Image"
5
  - "Krea-2"
6
  - "Qwen-Image-2.1"
7
  - "Fibo-2.0"
 
47
  model_type: "flux3"
48
  "MiniMax-H3-Image":
49
  model_type: "minimax-h3-image"
50
+ "P-H3-Image":
51
+ model_type: "p-h3-image"
52
  Krea-2:
53
  model_type: "krea-2"
54
  "SenseNova-U1.5":
yaml/model_defaults.yaml CHANGED
@@ -52,6 +52,14 @@ MiniMax-H3-Image:
52
  sampler_name: "euler"
53
  scheduler: "simple"
54
 
 
 
 
 
 
 
 
 
55
  Krea-2:
56
  _defaults:
57
  steps: 52
 
52
  sampler_name: "euler"
53
  scheduler: "simple"
54
 
55
+ P-H3-Image:
56
+ _defaults:
57
+ steps: 8
58
+ cfg: 1.0
59
+ sampler_name: "euler"
60
+ scheduler: "simple"
61
+ positive_prompt: "NOTE: P-H3-Image only supports the Txt2Img task."
62
+
63
  Krea-2:
64
  _defaults:
65
  steps: 52
yaml/model_list.yaml CHANGED
@@ -25,6 +25,14 @@ Checkpoint:
25
  unet: "minimax_h3_image_pruned_w4a8.safetensors"
26
  clip: "qwen3vl_32b_minimax_h3_nvfp4_awq.safetensors"
27
  vae: "minimax_h3_image_vae_fp16.safetensors"
 
 
 
 
 
 
 
 
28
  Krea-2:
29
  latent_type: latent
30
  models:
 
25
  unet: "minimax_h3_image_pruned_w4a8.safetensors"
26
  clip: "qwen3vl_32b_minimax_h3_nvfp4_awq.safetensors"
27
  vae: "minimax_h3_image_vae_fp16.safetensors"
28
+ P-H3-Image:
29
+ latent_type: none
30
+ models:
31
+ - display_name: "PrunaAI/P-H3-Image"
32
+ components:
33
+ unet: "p-h3-image_w4a8.safetensors"
34
+ clip: "p-h3-image_text_encoder_w4a8.safetensors"
35
+ vae: "p-h3-image_vae_int8_convrot.safetensors"
36
  Krea-2:
37
  latent_type: latent
38
  models: