Flex.1-alpha Guide: ComfyUI Setup and Local LoRA Training

Flex.1-alpha is ostris's downloadable 8B rectified-flow transformer for text-to-image generation. Run it through Diffusers or load its combined checkpoint in ComfyUI, then use the developer's AI-Toolkit configuration if you need LoRA training. Model card Training configuration

The model is released under Apache-2.0, with published weights and model-specific training instructions for local experimentation and adaptation. Model card

What are the key facts about Flex.1-alpha?

Use the full identifier ostris/Flex.1-alpha in a project record. The model name, checkpoint file, and training configuration should remain associated when moving an experiment between machines.

How does Flex.1-alpha support fine-tuning and LoRAs?

The developer describes Flex as a base model designed for fine-tuning. It has a trained guidance embedder that can be bypassed, allowing a specific training setup documented in AI-Toolkit. Card

The supplied configuration makes that setup concrete: it selects the Flex checkpoint, enables the FLUX architecture path, and bypasses guidance embedding during training. Official configuration

This is useful if your objective is learning a subject or visual treatment from a curated dataset. Begin by testing the base model so you can identify what the adaptation should change.

Write a target before collecting images. For a subject LoRA, that target might be a recognizable object across different surroundings; for a style experiment, it might be a consistent line treatment across unrelated subjects.

Keep those goals separate during review. An output that copies a training background may resemble the dataset without demonstrating the subject flexibility you wanted.

The developer's usage instructions also provide a combined ComfyUI checkpoint. That gives a documented route to try the base model before committing time to a training setup. Card

What limits and training settings apply to Flex.1-alpha?

The alpha model's documentation does not establish a universal inference latency or minimum VRAM requirement. Measure the workflow you actually intend to use, including loading and output review.

The training example includes quantization and gradient checkpointing. Treat it as an example configuration rather than proof that every dataset and setting fits the same device. Training configuration

The model card lists a 512-token input length. Put the indispensable subject and composition instructions first instead of relying on an increasingly long collection of style words. Card

Training and inference require separate reasoning. A setting used to optimize the model's learning process is not automatically the setting you should expose in an image-generation interface.

In particular, the sample configuration explicitly calls for bypassing the guidance embedder during training. Preserve that model-specific instruction when adapting the example. Training configuration

A successful training run is also not enough to judge a LoRA. Reserve prompts and situations that are different from the training captions, then inspect whether the adaptation generalizes to them.

How do you run Flex.1-alpha in ComfyUI or Diffusers?

For a first ComfyUI run, download the combined Flex.1-alpha.safetensors file from the official model repository. The card says to place it in the checkpoints folder and use it as a FLUX.1-dev-style checkpoint. Card

In a standard ComfyUI layout, that means ComfyUI/models/checkpoints/. Follow the model card's combined-file route consistently rather than mixing it with instructions for separately loaded components.

For the surrounding interface, use the ComfyUI pillar. Keep a saved workflow for the unchanged model before introducing an adapter.

Diffusers is another documented access path. After installing the dependencies listed on the model page, this minimal example loads the official repository and saves an image. Model usage

import torch
from diffusers import DiffusionPipeline

pipe = DiffusionPipeline.from_pretrained(
    "ostris/Flex.1-alpha",
    torch_dtype=torch.bfloat16,
)
pipe.to("cuda")
image = pipe(
    "A wooden toy boat on a workbench, soft daylight"
).images[0]
image.save("flex-baseline.png")

This example assumes a suitable CUDA environment and enough memory for the chosen configuration. It is a setup illustration, not a measured hardware recommendation.

For LoRA training, open the developer's linked AI-Toolkit example. It documents image folders, matching text captions, the checkpoint identifier, and sampling settings. Configuration

The example pairs an image with a text file sharing its base name. Prepare that pairing carefully so a caption describes the actual image rather than a neighboring file. Dataset configuration

Then replace the example dataset path and output name with project-specific values. Review the sample prompts too: choose prompts that reveal the particular subject or style you want to learn.

Keep the baseline outputs, training configuration, and saved adapter together. When comparing checkpoints, use the same held-out prompts and record what changes at each stage.

The LoRA guide provides the next workflow context. Verify the adapter's intended base model before interpreting a poor result as a training failure.

How does Flex.1-alpha compare with FLUX.1-schnell?

FLUX.1-schnell is a relevant alternative because it sits in Flex's documented lineage. Black Forest Labs describes schnell as a 12B model trained for generation in one to four inference steps. FLUX card

Do not transfer schnell's step count into a Flex performance claim. Choose the documented baseline for each model, then compare accepted output quality and total effort on your own brief.

What else should you know before using Flex.1-alpha?

Is Flex.1-alpha open weight?

ostris publishes Flex.1-alpha's weights with an Apache-2.0 license declaration. Obtain them from the official repository so the file and its model documentation stay connected. Card

Does Flex.1-alpha support LoRA training?

Flex.1-alpha has a dedicated AI-Toolkit LoRA configuration. Its training section explicitly bypasses the guidance embedder, so use that model-specific example as the starting point. Configuration

Where does its combined ComfyUI file go?

The Flex.1-alpha model card places Flex.1-alpha.safetensors in ComfyUI's checkpoints folder. Use that combined-file workflow with a checkpoint loader. Card

Is Flex guaranteed to be faster than FLUX.1-schnell?

The Flex.1-alpha and FLUX.1-schnell model cards do not establish a fixed speed advantage for Flex.1-alpha. Benchmark both documented workflows on your hardware and include the number of attempts needed for an accepted image. Flex card FLUX card

Sources

• ostris Flex.1-alpha model card
• Official AI-Toolkit Flex LoRA configuration
• Black Forest Labs FLUX.1-schnell model card

Related guides on PromptZone

• Best SDXL Models in 2026
• ComfyUI 2026: The Complete Guide
• AI Model Releases Timeline