ImageGen / chain_injectors /moss_image_edit_injector.py
RioShiina's picture
feat: integrate MOSS-Image Edit with reference images and LoRA
7481b8b verified
Raw History Blame Contribute Delete
10.3 kB
import os
try:
from utils.app_utils import ensure_file_downloaded
except ImportError:
ensure_file_downloaded = None
def create_node(assembler, class_type, title=None):
try:
node = assembler._get_node_template_from_api(class_type)
except Exception:
try:
node = assembler._get_node_template(class_type)
except Exception:
node = {
"inputs": {},
"class_type": class_type,
"_meta": {"title": title or class_type}
}
if title:
node.setdefault('_meta', {})['title'] = title
return node
def inject(assembler, chain_definition, chain_items):
"""
Injects MOSS-Image Edit (Reference Image Edit) into the workflow:
1. Loads the officially packaged Edit LoRA (moss_image_edit_lora_rank_64_bf16.safetensors)
onto the model path.
2. Loads and scales up to 10 reference images with ImageScaleToTotalPixels.
3. Replaces positive/negative prompt encoding with native TextEncodeMOSSImageEdit nodes,
which attach reference image latents and text tokens.
4. Connects patched model and conditioned positive/negative outputs to KSampler.
"""
if not chain_items:
return
# Normalize chain_items: dict or list of dicts/strings
edit_format = "Moss-Image-Edit"
strength = 1.0
valid_images = []
if isinstance(chain_items, dict):
chain_items = [chain_items]
if isinstance(chain_items, list):
for item in chain_items:
if not item:
continue
if isinstance(item, dict):
raw_imgs = item.get('images')
if isinstance(raw_imgs, list):
for img in raw_imgs:
p = img.get('image') or img.get('filename') or img.get('path') if isinstance(img, dict) else img
if p and isinstance(p, str) and p.strip():
valid_images.append(p.strip())
elif isinstance(raw_imgs, str) and raw_imgs.strip():
valid_images.append(raw_imgs.strip())
p = item.get('image') or item.get('filename') or item.get('path')
if p and isinstance(p, str) and p.strip():
valid_images.append(p.strip())
if 'edit_format' in item:
edit_format = item['edit_format']
if 'strength' in item:
try:
strength = float(item['strength'])
except (ValueError, TypeError):
pass
elif isinstance(item, str) and item.strip():
valid_images.append(item.strip())
if not valid_images:
return
valid_images = valid_images[:10]
lora_filename = "moss_image_edit_lora_rank_64_bf16.safetensors"
if ensure_file_downloaded:
try:
ensure_file_downloaded(lora_filename)
except Exception as e:
print(f"Warning: Failed to ensure '{lora_filename}' downloaded: {e}")
ksampler_name = chain_definition.get('ksampler_node', 'ksampler')
pos_prompt_name = chain_definition.get('pos_prompt_node', 'pos_prompt')
neg_prompt_name = chain_definition.get('neg_prompt_node', 'neg_prompt')
clip_loader_name = chain_definition.get('clip_loader_node', 'clip_loader')
vae_loader_name = chain_definition.get('vae_loader_node', 'vae_loader')
if ksampler_name not in assembler.node_map:
print(f"Warning: Target node '{ksampler_name}' for MOSS-Image Edit chain not found. Skipping.")
return
ksampler_id = assembler.node_map[ksampler_name]
if 'model' not in assembler.workflow[ksampler_id]['inputs']:
print(f"Warning: KSampler node '{ksampler_name}' is missing 'model' input. Skipping.")
return
current_model_connection = assembler.workflow[ksampler_id]['inputs']['model']
vae_connection = None
if vae_loader_name in assembler.node_map:
vae_connection = [assembler.node_map[vae_loader_name], 0]
else:
for node_id, node in assembler.workflow.items():
if isinstance(node, dict) and node.get('class_type') == 'VAELoader':
vae_connection = [node_id, 0]
break
if not vae_connection:
print("Warning: VAE connection not found for MOSS-Image Edit. Skipping.")
return
pos_prompt_id = assembler.node_map.get(pos_prompt_name)
neg_prompt_id = assembler.node_map.get(neg_prompt_name)
clip_connection = None
if clip_loader_name in assembler.node_map:
clip_connection = [assembler.node_map[clip_loader_name], 0]
elif pos_prompt_id and 'clip' in assembler.workflow.get(pos_prompt_id, {}).get('inputs', {}):
clip_connection = assembler.workflow[pos_prompt_id]['inputs']['clip']
else:
for node_id, node in assembler.workflow.items():
if isinstance(node, dict) and node.get('class_type') == 'CLIPLoader':
clip_connection = [node_id, 0]
break
if not clip_connection:
print("Warning: CLIP connection not found for MOSS-Image Edit. Skipping.")
return
# 1. Inject LoRA loader on model path if strength > 0
if strength > 0.0:
lora_loader_id = assembler._get_unique_id()
lora_loader_node = create_node(assembler, "LoraLoaderModelOnly", "Load LoRA (MOSS-Image Edit)")
lora_loader_node['inputs']['lora_name'] = lora_filename
lora_loader_node['inputs']['strength_model'] = strength
lora_loader_node['inputs']['model'] = current_model_connection
assembler.workflow[lora_loader_id] = lora_loader_node
current_model_connection = [lora_loader_id, 0]
megapixels = float(chain_definition.get('megapixels', 1.0))
# 2. Process Reference Images
scale_image_ids = []
for i, img_filename in enumerate(valid_images):
load_id = assembler._get_unique_id()
load_node = create_node(assembler, "LoadImage", f"Load Reference Image {i+1} (MOSS Edit)")
load_node['inputs']['image'] = img_filename
assembler.workflow[load_id] = load_node
scale_id = assembler._get_unique_id()
scale_node = create_node(assembler, "ImageScaleToTotalPixels", f"Scale Reference {i+1} (MOSS Edit)")
scale_node['inputs']['upscale_method'] = "lanczos"
scale_node['inputs']['megapixels'] = megapixels
scale_node['inputs']['resolution_steps'] = 1
scale_node['inputs']['image'] = [load_id, 0]
assembler.workflow[scale_id] = scale_node
scale_image_ids.append(scale_id)
# 3. Extract Prompt Texts
pos_text = ""
if pos_prompt_id and pos_prompt_id in assembler.workflow:
pos_text = assembler.workflow[pos_prompt_id]['inputs'].get('text', '') or assembler.workflow[pos_prompt_id]['inputs'].get('prompt', '')
elif hasattr(assembler, 'ui_values') and isinstance(assembler.ui_values, dict):
pos_text = assembler.ui_values.get('positive_prompt') or assembler.ui_values.get('prompt') or ''
neg_text = ""
if neg_prompt_id and neg_prompt_id in assembler.workflow:
neg_text = assembler.workflow[neg_prompt_id]['inputs'].get('text', '') or assembler.workflow[neg_prompt_id]['inputs'].get('prompt', '')
elif hasattr(assembler, 'ui_values') and isinstance(assembler.ui_values, dict):
neg_text = assembler.ui_values.get('negative_prompt') or assembler.ui_values.get('neg_prompt') or ''
# 4. Positive TextEncodeMOSSImageEdit
pos_edit_id = assembler._get_unique_id()
pos_edit_node = create_node(assembler, "TextEncodeMOSSImageEdit", "Text Encode MOSS-Image Edit")
pos_edit_node['inputs']['clip'] = clip_connection
pos_edit_node['inputs']['prompt'] = pos_text
pos_edit_node['inputs']['vae'] = vae_connection
pos_edit_node['inputs']['edit_format'] = edit_format
pos_edit_node['inputs']['image_strength'] = 1.0
for i, sc_id in enumerate(scale_image_ids):
pos_edit_node['inputs'][f"images.image_{i}"] = [sc_id, 0]
pos_edit_node['inputs'] = {k: v for k, v in pos_edit_node['inputs'].items() if v is not None}
assembler.workflow[pos_edit_id] = pos_edit_node
# 5. Negative TextEncodeMOSSImageEdit
neg_edit_id = assembler._get_unique_id()
neg_edit_node = create_node(assembler, "TextEncodeMOSSImageEdit", "Text Encode MOSS-Image Edit")
neg_edit_node['inputs']['clip'] = clip_connection
neg_edit_node['inputs']['prompt'] = neg_text
neg_edit_node['inputs']['vae'] = vae_connection
neg_edit_node['inputs']['edit_format'] = edit_format
neg_edit_node['inputs']['image_strength'] = 1.0
for i, sc_id in enumerate(scale_image_ids):
neg_edit_node['inputs'][f"images.image_{i}"] = [sc_id, 0]
neg_edit_node['inputs'] = {k: v for k, v in neg_edit_node['inputs'].items() if v is not None}
assembler.workflow[neg_edit_id] = neg_edit_node
# 6. Reconnect model and redirect conditioning references in workflow
assembler.workflow[ksampler_id]['inputs']['model'] = current_model_connection
# Redirect all existing connections that referenced pos_prompt_id / neg_prompt_id
if pos_prompt_id:
for node in assembler.workflow.values():
if isinstance(node, dict) and 'inputs' in node:
for k, v in node['inputs'].items():
if isinstance(v, list) and len(v) == 2 and v[0] == pos_prompt_id:
node['inputs'][k] = [pos_edit_id, 0]
if neg_prompt_id:
for node in assembler.workflow.values():
if isinstance(node, dict) and 'inputs' in node:
for k, v in node['inputs'].items():
if isinstance(v, list) and len(v) == 2 and v[0] == neg_prompt_id:
node['inputs'][k] = [neg_edit_id, 0]
# Clean up original CLIPTextEncode nodes if they are in workflow
if pos_prompt_id and pos_prompt_id in assembler.workflow:
del assembler.workflow[pos_prompt_id]
if neg_prompt_id and neg_prompt_id in assembler.workflow:
del assembler.workflow[neg_prompt_id]
print(f"MOSS-Image Edit injector applied with {len(valid_images)} reference image(s) (format: {edit_format}, strength: {strength}).")