SDXL Deprecation & Migrating Your Models to Flux

Scenario is deprecating SDXL as a base model for custom model training. This article explains what this means, what happens to your existing SDXL models, and how to retrain them on Flux for better results.

What Is Changing and Why

SDXL (Stable Diffusion XL) has been the foundation for many custom-trained models on Scenario since the platform launched. Flux — developed by Black Forest Labs — now significantly outperforms it across all key dimensions: prompt adherence, image quality, trainability from smaller datasets, and generation speed at lower cost.

SDXL is now considered a legacy base model and is no longer the recommended option for new training jobs.

What Happens to My Existing SDXL Models

Your existing SDXL-trained models will not be deleted. You can continue using them for generation as normal. However, new training jobs default to Flux, and SDXL models cannot be merged with Flux LoRAs — model merging only works within the same base model family. As Scenario continues to invest in Flux-based features and optimizations, the capability gap between the two will widen over time.

How to Retrain Your Model on Flux

Migrating is a retraining process. You use the same dataset from your SDXL model and train a new model on the Flux base.

Go to Your Models, open your SDXL model, and click Training Set. Download your original training images from there.

Step 2 — Create a new model

Go to Train > New Model and select the appropriate model type (Style, Character, or Subject).

Step 3 — Select Flux as the base model

In the base model selector, choose Flux 2 Dev for highest quality or Flux 2 Klein for faster, lower-cost generation. Do not select SDXL.

Step 4 — Upload your dataset

Upload the same images used for your SDXL model. Flux performs well with 10 to 25 images. If your original dataset was larger, prioritize the highest-quality and most representative images.

Step 5 — Review captions

Scenario auto-captions your images after upload. Review them carefully — Flux responds well to detailed, descriptive captions that include style descriptors, lighting conditions, and subject specifics.

Step 6 — Train and compare epochs

Set test prompts that match your typical use case. After training, compare epochs and select the one that best balances style fidelity with prompt flexibility.

Key Differences When Using Your New Flux Model

Once your Flux model is trained, a few settings behave differently compared to your SDXL model.

The recommended Guidance range changes from 6–12 (SDXL) to 3–5 (Flux). Sampling Steps drop from around 30 to around 28. Prompt style also shifts: SDXL responded better to short, keyword-heavy prompts, while Flux benefits from detailed, sentence-based descriptions. Caption style follows the same pattern — brief labels worked for SDXL, while full descriptive captions produce better results with Flux.

If your Flux model feels less controlled than your SDXL model was, try increasing Guidance slightly to 5–6 and adding more specific style keywords to your prompts. Flux is more responsive to prompt detail.

Flux Kontext: An Alternative for Editing-Focused Models

If your SDXL model was primarily used for character editing, style transfer, or instruction-based modifications, consider training on Flux Kontext instead of standard Flux 2. Flux Kontext is built for instruction-based transformations and uses pairs of before-and-after images as training data, making it ideal for workflows where you want to apply a consistent edit or transformation style. See Train a Custom Flux Kontext LoRA for a full guide.

Frequently Asked Questions

I have many SDXL models. Do I have to retrain all of them right now?

No. Prioritize the models you actively use in production. Models used infrequently can stay as-is until you have a natural reason to refresh them.

Can I import an SDXL LoRA from outside Scenario and convert it to Flux?

No. LoRAs are architecture-specific and are not cross-compatible between SDXL and Flux.

Will my SDXL model's generation quality get worse over time?

No. SDXL inference quality stays the same. What changes is that new platform features and improvements are built around Flux, so the gap in capability will widen over time.