Use normalized LoRA stacking to prevent character bleed in SDXL ComfyUI workflows.

DeepSeek-V3 has become the main choice for writing the JSON logic and custom node expressions used to stabilize SDXL LoRA blending in ComfyUI. Its preference over Claude 3.5 Sonnet comes from Sonnet’s generally conservative weight recommendations. When several LoRAs cause “character bleed” or weaken the likeness of a particular subject, the interaction between the CLIP text encoder and the UNet weights is almost always responsible.

Chaining LoRAs frequently produces messy pixels. A common approach connects two LoRAs and estimates their weights by trial, but benchmark comparisons distinguish a “Weighted Blend” method from “Sequential Application.” With sequential application, LoRA A -> LoRA B -> KSampler creates a cumulative effect in which the second LoRA can overpower the first. That problem becomes more likely when the first LoRA contains a strong character trigger.

A “LoRA Stack” normalizes the combined weight to 1.0, which supports more consistent results. A character LoRA can use 0.6 while its accompanying style LoRA uses 0.4. A combined weight beyond 1.2 usually produces the metallic sheen or contrast clipping often seen in SDXL.

Prompt weighting also stabilizes character identity. A strength modifier preserves the character while the LoRA supplies its visual characteristics, rather than using only the trigger word. The same prompt structure works across different samplers:

(masterpiece:1.2), (photorealistic:1.1), <lora:Char_Identity:0.7> [character_trigger_word:1.3], <lora:Style_Blend:0.3> cinematic lighting, highly detailed skin pores

Gemini 1.5 Pro demonstrates a strong understanding of spatial relationships in ComfyUI regional prompting. When the task is limited to optimizing weights, DeepSeek-V3 provides more exact numerical recommendations by drawing on training epoch data.

Sequential loading and normalized stacking involve different compromises. Sequential LoRA Loading is fast to configure and works well for uncomplicated style overlays, but its heavy “over-baking” effect allows the final LoRA in the chain to dominate the image. Normalized Weight Stacking offers noticeably stronger character fidelity and reduces the chance of a generic “AI look.” That improvement requires manual adjustment for every new LoRA combination and slows iteration.

Sampler adjustments can resolve changes in facial appearance. If adding a clothing or background LoRA alters the character’s face, connect the character LoRA after the style LoRA. Although this ordering can appear counterintuitive, the final applied weight has disproportionate influence over SDXL’s final latent space. Using dpmpp_2m_sde with karras is another recommended adjustment. Compared with euler_a, this pairing handles conflicting gradients from two LoRAs more smoothly and reduces artifacts around character edges.