Have you trained a dedicated portrait LoRA but can only generate close-up headshots? When you excitedly connect the LoRA to ComfyUI, only to find that it generates nothing but “ID photo-style” close-ups, and you can’t get half-body/full-body compositions to work—this is a pitfall that 90% of AI portrait players have encountered.
Today, we will guide you step by step from the training framework origins to node flow configuration, teaching you how to break through the “close-up curse”!
1.Problem Origin: Why Does Your LoRA Only Generate Close-Ups?
1. Analysis of FluxGYM Training Characteristics
This framework defaults to using the progressive resolution training method:
•In the early stages, it learns facial details at a resolution of 512×512•Later, it gradually increases the resolution to enhance local features⚠️ Side effect: The model will form a path dependency of “close-up = high quality”
2. ComfyUI Loading Mechanism
Unlike WebUI’s automatic merging, ComfyUI’s explicit node connections require:
•You must manually balance <span>Prompt weight</span>, <span>LoRA strength</span>, and <span>ControlNet guidance</span> three forces
2.ComfyUI Breakthrough Solutions
📍 Core Idea
“Trinity Control Method”:
[Composition Prompt] ← Counteract → [LoRA Characteristics] + [ControlNet Strong Guidance]🔧 Practical Configuration
1. Fine-tuning the LoraLoader Node
# Key parameter settingsmodel = "your_model.safetensors" strength = 0.65 # FluxGYM training suggests a range of 0.6-0.75clip_strength = 0.8 # Strengthen the text encoder separately❗ Pitfall Reminder: The LoRA trained with FluxGYM is more sensitive to clip_strength, it is recommended to be higher than that of regular models
2. Dual ControlNet Combination
| ControlNet Type | Function | Recommended Model |
| OpenPose | Locks body proportions/poses | thibaud_controlnet |
| Depth | Controls depth of field to prevent the face from being overly prominent | depth_zoe |
# Node parameter example"preprocessor": "none" # Directly input pose image "control_net_weight": [0.85, 0.75] # OpenPose weight > Depth "start_percent": 0.15 # Start intervention at step 15 3. Prompt Engineering Template
Positive Prompt = """(full-body portrait:1.4), (perfect proportions:1.3), [your exclusive trigger word], (detailed clothing:1.2), (standing in city street:1.1), (natural lighting),"""Negative Prompt = """(close-up:1.6), (floating head:1.4), (deformed hands:1.3), (duplicate limbs:1.3), (low resolution:1.2)💡 FluxGYM Special Reminder: The LoRA trained with this framework is more sensitive to <span>(close-up)</span> negative prompts
3.Advanced Training Techniques
1. Training Data Optimization
•Add dynamic cropping strategy in the FluxGYM configuration:
# flux_train.yamldata_processing:random_crop: min_scale: 0.6 # Minimum cropping ratio max_scale: 1.0 target_size: 768 # Match output resolution2. Layered Tagging Method
Use dual tag labeling for full-body training images:
[Your Character Name], (full-body:1.2), (standing:0.9), (studio lighting:0.7), (high-end fashion)4.Common Troubleshooting
| Phenomenon | Solution |
| Face normal, body deformed | Enable Hi-Res Fix (2x super resolution) + Increase Depth control net weight |
| Clothing style unstable | Add clothing close-up images to the LoRA training set + Add prompt <span>(consistent style:1.2)</span> |
| Generated results too realistic | Adjust CFG Scale to the range of 6-7 + Add <span>(cinematic lighting:1.3)</span> |
5.Ultimate Solution: Training-Generation Joint Optimization
•Automatically identify composition types during the training phase•Intelligently match corresponding fine-tuning models during generation
Technical Summary: Through the combination of “Prompt Directional Guidance + ControlNet Hard Constraints + LoRA Weight Softening”, even a FluxGYM training model focused on the face can output stunning full-body compositions in ComfyUI. It is recommended to save the node flow diagram in this article as a configuration baseline, and gradually try dynamic enhancement of training data in the future.
Next Issue Preview: “ComfyUI Multi-Person Pose Control: How to Make LoRA Characters Interact Naturally?”
Scan to follow the XYAIXY public account
#AI Painting #ComfyUI Black Technology #LoRA Training #XYAI Laboratory