LoRAs from the artist's perspective: a working playbook

ComfyUIComfyUIThe dominant node-based interface for running diffusion models locally. Started by 'comfyanonymous' in January 2023, now stewarded by Comfy Org. and most other tools let you load multiple LoRAs simultaneously and stack their effects. This is enormously powerful and almost always how serious creative work happens. A typical professional workflow stacks: a style LoRAStyle LoRAA LoRA trained on images that share an aesthetic but vary in subject, so it can render any prompt in that art style, photographic look, or brand identity. (the look), a character LoRACharacter LoRAA small model trained on images of one specific subject so that a fixed identity can be summoned consistently across an entire production. (the subject), a quality enhancer LoRALoRA (Low-Rank Adaptation)A fine-tuning technique that adapts a base model to a specific style or concept using a small additional file. The standard way to customize open-source models. (a detail or skin texture upgrade), and sometimes a concept LoRAConcept LoRAA LoRA that teaches or repairs a specific concept, action, or object the base model handles badly, such as correct hands or a particular pose. (to fix a known failure mode of the base model). The four LoRAs combined produce something none of them could produce individually.

The catch is that LoRAs interact. When you stack two LoRAs at full strength, they often fight, each one is trying to push the model in its own direction, and the result is muddy. The technique is to lower the strength of each LoRA proportionally as you stack more of them. If a single LoRA looks right at strength 1.0, two LoRAs together usually look right at 0.7 each, three LoRAs together at 0.5 each. There is no formula, you tune by eye. The art is in finding the combination of LoRAs and strengths that produces the look you want without any one LoRA dominating.

Some pairs of LoRAs are mutually compatible (a style LoRA and a subject LoRASubject LoRAA LoRA trained to reproduce one specific person, character, or object across new scenes and poses, the most common and easiest kind to train. from different training sets usually combine well). Some pairs are mutually exclusive (two style LoRAs trained on contradictory aesthetics will produce mush). Learning which LoRAs combine well is a matter of experience and is a real source of competitive advantage for working AI artists. The artists who consistently produce great work tend to have personal libraries of 20 to 50 LoRAs they have tested in combination and know exactly how to blend.

The strength parameterParameterOne individual weight. Model size is quoted in parameters, in the millions or billions. More parameters can mean more capability and more cost. on a LoRALoRA (Low-Rank Adaptation)A fine-tuning technique that adapts a base model to a specific style or concept using a small additional file. The standard way to customize open-source models. is not a binary on/off switch. It is a continuous creative dial. At strength 0.2, a style LoRAStyle LoRAA LoRA trained on images that share an aesthetic but vary in subject, so it can render any prompt in that art style, photographic look, or brand identity. contributes a faint suggestion of its aesthetic, colors, mood, lighting bias. At strength 0.5, the LoRA is clearly present but the base model's general capabilities are still dominant. At strength 0.8, the LoRA is the primary aesthetic driver. At strength 1.0, the LoRA is fully applied as trained. Above 1.0, you are pushing harder than the LoRA was designed for.

Different strengths are appropriate for different creative goals. For a 'in the style of' promptPromptThe text description you provide to a model to specify what you want it to generate. where you want the style to be obvious, run the style LoRA at 0.7 to 0.9. For a subtle stylistic nudge that will read as the model's natural output, run it at 0.3 to 0.5. For a character that needs to be unmistakably the same character across many generations, run the subject LoRASubject LoRAA LoRA trained to reproduce one specific person, character, or object across new scenes and poses, the most common and easiest kind to train. at 0.8 to 1.0. For a character that should feel like the same person but with room for interpretation, run it at 0.5 to 0.7. The dial is part of your creative vocabulary.

Many LoRAs were trained with a specific trigger wordTrigger wordA unique, deliberately unusual token used in training captions to name what a LoRA teaches, so that including the word in a prompt activates the learned subject or style. that is supposed to appear in your promptPromptThe text description you provide to a model to specify what you want it to generate.. The convention is that the LoRALoRA (Low-Rank Adaptation)A fine-tuning technique that adapts a base model to a specific style or concept using a small additional file. The standard way to customize open-source models. only takes effect when the trigger word is present, and is dormant otherwise. This is partly a useful feature and partly an artifact of how training conventions evolved. In practice, modern LoRAs in FLUX and SDXL are often trained without strict trigger words, they just modify the base model's behavior whenever they are loaded, and you can use them with any prompt.

The practical guidance is: if the LoRA's documentation specifies a trigger word, use it. If it does not, just load the LoRA and prompt normally. If you are getting weak results from a LoRA you expect to be strong, check whether it has a trigger word you forgot to include.

One of the hardest problems in working with image models is producing many different images of the same character that all clearly show the same person. Even with a good subject LoRASubject LoRAA LoRA trained to reproduce one specific person, character, or object across new scenes and poses, the most common and easiest kind to train., the model has some natural variation that pushes each generation slightly toward generic-person territory. Professional workflows for character consistencyCharacter consistencyThe problem of generating the same character, with the same face and identity, across many images, poses, and scenes. It is the main bottleneck holding back AI narrative work. typically combine multiple techniques: a strong subject LoRA, a fixed seed (so the random component of the generation is the same across runs), an Intellectual Property (IP)-Adapter (a different mechanism that uses a reference imageReference imageAn image fed to a model as conditioning to carry a subject, style, or composition into new output, rather than describing it in words. to bias the generation), and sometimes ControlNetControlNetA conditioning technique that lets you provide structural guidance (depth, edges, pose) to a diffusion model to control spatial layout. with a face landmark map (to enforce the same facial structure).

The most reliable workflow for character consistency in 2026 is: train a subject LoRA on 25-40 high-variation reference images, generate variations with a fixed seed and a careful promptPromptThe text description you provide to a model to specify what you want it to generate. template, run the outputs through a face-swap model as a final pass to enforce identity if needed, and accept that 1 in 5 generations will need to be discarded. Even with all of these techniques, some drift is unavoidable, the technology is good enough for editorial work and concept art but not yet good enough to produce thousands of perfectly consistent shots without human review.