What is Step1X-Edit-v1p2?
Step1X-Edit-v1p2 is StepFun’s reasoning-enhanced image-to-image editing model. It takes a reference image together with a text instruction, then generates an edited image that attempts to follow the requested change. For example, a user could ask the system to add an object, remove an object, or change an existing visual attribute in the supplied image.
The model is published as open weights on Hugging Face under the Apache 2.0 license. The canonical repository is stepfun-ai/Step1X-Edit-v1p2. In StepFun’s associated technical material, this release is also referred to as ReasonEdit-S. It belongs to the Step1X-Edit family and is distinct from the earlier Step1X-Edit, Step1X-Edit-v1p1, and Step1X-Edit-v1p2-preview checkpoints.
Rather than being a general-purpose conversational model, Step1X-Edit-v1p2 is specialized for visual transformation. It does not target chat, speech, transcription, embeddings, video generation, or ordinary text generation.
How the reasoning and reflection pipeline works
The defining feature of v1p2 is that reasoning is integrated into the editing workflow. The pipeline can run with optional thinking and reflection modes. Thinking can analyze and reformulate the user’s instruction before image generation. This is useful when a request contains several changes, ambiguous relationships, or detailed requirements that may benefit from a more explicit intermediate description.
Reflection adds another evaluation stage. The system can assess intermediate results and select a better final image. This does not mean the model becomes a general reasoning or chat system: its reasoning is focused on interpreting and evaluating image-editing tasks.
The official usage pattern uses a custom Step1XEditPipelineV1P2 implementation. The pipeline accepts an input image, a text prompt, inference settings, and optional thinking or reflection flags. When those features are enabled, the pipeline can return additional reasoning-related information, including a reformulated prompt and reflection metadata, alongside the generated image.
Inputs, outputs, and supported modalities
| Capability | Support |
|---|---|
| Text input | Yes, as an editing instruction |
| Image input | Yes, as a reference image |
| Image output | Yes, the edited image |
| Audio input or output | No documented support |
| Video input or output | No documented support |
| Text generation | Not the model’s primary output; optional prompt and reasoning metadata may be returned |
The essential workflow is therefore image plus instruction to edited image. The model’s multimodal nature comes from combining visual input with text, not from supporting every media type. There is no supplied specification for a context window or maximum output-token limit; those language-model-style fields are not applicable or documented for this image editing checkpoint.
Main strengths
Handling complex edit instructions
Step1X-Edit-v1p2 is most differentiated when an edit requires more than a simple visual change. Thinking mode can turn a loosely phrased request into a more explicit editing plan, while reflection can provide an additional quality-control step. These mechanisms are intended to improve instruction following and editing accuracy, particularly when several constraints need to be respected at once.
Open-weight deployment
The Apache 2.0 release allows developers and researchers to obtain the weights and run the system themselves, subject to the license terms and their own infrastructure requirements. Local execution can be preferable when an organization needs control over its environment, wants to experiment with the pipeline, or does not want to depend on a per-request hosted image API.
Reported benchmark results
StepFun reports a GEdit-Bench overall score of 7.58 for the v1p2 configuration with thinking and reflection enabled, and a KRIS-Bench score of 60.93. These are provider-reported results, not independent guarantees. Results can depend on the benchmark version, prompts, inference configuration, and evaluation process, so they should be used as reference points rather than predictions for every image.
Deployment requirements and practical cost
The official instructions use PyTorch, Transformers 4.55.0, a development branch of Diffusers, and bfloat16 weights on CUDA-capable hardware. RegionE is available as an optional acceleration component. Because the model is intended for local inference, successful setup requires more than installing a standard consumer image-generation application: users must align the model code, library versions, hardware, and CUDA environment.
StepFun recommends a GPU with approximately 80 GB of memory for efficient, high-quality generation. The broader Step1X-Edit project documents memory-saving approaches and quantized or offloaded configurations, but these may involve compromises in speed, convenience, or output quality. The supplied research does not establish one universal minimum hardware configuration for v1p2, so the 80 GB figure should be treated as a practical recommendation rather than a guaranteed hard requirement.
No official StepFun per-request, per-image, or per-token hosted API price was identified for this exact checkpoint. The weights are available under Apache 2.0, but open weights are not free to operate: users still pay for GPU hardware, cloud instances, electricity, storage, and engineering time. For small workloads, a hosted image-editing service may be simpler or cheaper overall. For sustained workloads or research use, self-hosting may offer better control and predictable infrastructure economics.
Reasoning, speed, and capability trade-offs
Thinking and reflection can improve the handling of difficult instructions, but they add processing steps compared with a basic image-editing pass. Users who need the fastest possible turnaround may prefer a simpler configuration, while users editing important or complex images may accept additional latency for the extra interpretation and evaluation stages.
The available editorial assessment rates the model’s reasoning capability at 8 out of 10, speed at 4 out of 10, and cost at 8 out of 10. These are editorial scores, not StepFun specifications or benchmark results. They reflect the practical trade-off suggested by the model’s design: strong task-focused reasoning and open-weight cost flexibility, balanced against demanding local hardware and a comparatively complex deployment process.
There is no documented tool or function-calling interface, streaming API, batch API, or JSON-output mode for this checkpoint. It should therefore be treated as an image-generation pipeline rather than an agent platform or structured-output service.
Best use cases
- Instruction-based image editing: adding, removing, or modifying objects and attributes in an existing image.
- Complex visual modifications: requests with multiple changes or detailed spatial and semantic constraints.
- Research and experimentation: testing reasoning-enhanced image editing with accessible model weights.
- Private or controlled workflows: running inference in infrastructure managed by the user instead of sending images to a third-party hosted service.
- Repeated local generation: workloads where owning or reserving suitable GPU capacity is justified.
When to choose Step1X-Edit-v1p2
Choose Step1X-Edit-v1p2 when the primary task is editing an existing image, open-weight access matters, and you can support a demanding CUDA-based deployment. It is especially suitable when instruction interpretation is more important than minimum latency and when thinking or reflection can help resolve a complicated edit.
A hosted image-editing option may be more appropriate when you need a simple web or API workflow, do not have access to a large GPU, or only edit images occasionally. A lighter image editor may also be preferable for straightforward changes where the additional reasoning stages do not justify their latency and infrastructure cost. Conversely, a general multimodal or conversational model would be a better fit for dialogue, document analysis, coding, tool use, or broad text-and-image tasks; those are outside Step1X-Edit-v1p2’s documented purpose.
Limitations and availability
Step1X-Edit-v1p2 has no identified official hosted pricing or documented first-party API service for this exact checkpoint. It requires specialized dependencies, CUDA-compatible hardware, and substantial memory for efficient use. The model also has no documented support for audio, video, speech, embeddings, tool calling, or general-purpose text generation.
Benchmark scores should not be interpreted as guarantees across different prompts, languages, image types, resolutions, or hardware configurations. The model’s release information identifies it as a current open-weight checkpoint dated November 26, 2025 in the supplied research, while its long-term maintenance and deprecation status are not documented. For deployment, users should consult the official model card and repository for the version-specific installation instructions rather than assuming compatibility with a generic Diffusers installation.