Ovis Image: Precise Text-to-Image Generation Playground & API | RunComfy

Model overview

Provider

AIDC-AI (Alibaba)

Task

Text-to-image generation

Architecture

Advanced multimodal architecture optimized for high-level semantic understanding and visual generation

Resolution

Supports multiple aspect ratios including Square HD, Portrait, and Landscape formats

Key strengths

• Superior Prompt Adherence: Deep understanding of complex textual descriptions and spatial relationships.
• Photorealistic Quality: Excellent handling of lighting, textures, and material properties.
• Versatile Styles: Capable of generating everything from "gorpcore" photography to artistic illustrations.
• Efficient Inference: Optimized for rapid generation without sacrificing image fidelity.

Ovis Image represents the latest advancements from AIDC-AI, leveraging deep visual-language alignment to ensure that the generated output strictly follows the user's intent. Unlike older diffusion models that may struggle with long prompts, Ovis Image maintains coherence across detailed scenarios.

How Ovis Image runs on RunComfy

On RunComfy, Ovis Image is hosted as a managed, scalable service exposed in three complementary ways:

Playground UI

Prompt, adjust parameters like guidance scale and steps, and run text-to-image jobs directly in your browser.

Ideal for testing prompt fidelity and exploring the capabilities of Ovis Image before integration.

Playground API

From the playground view, you can use the model as an API and call it from your own apps or services.

This provides a private, production-ready endpoint matching the configuration you tested.

In all cases, inference runs on RunComfy’s cloud GPUs—no local hardware, drivers, or downloads needed.

Input parameters

Ovis Image on RunComfy exposes a streamlined set of parameters designed for ease of use and consistent results.

Core text and guidance

When configuring the model, the most critical parameter is the prompt. Ovis Image is specifically designed to handle long, descriptive string inputs (such as specific camera angles, lighting, or outfit details) with high proficiency. Conversely, you can use the negative_prompt string to instruct Ovis Image on what to exclude, such as "blur," "low quality," or "distortion."

For processing control, Ovis Image utilizes num_inference_steps, which defines the number of denoising steps. While the default is 28, Ovis Image typically operates within a range of 20 to 50 steps; higher values increase detail but require more processing time. Furthermore, the guidance_scale (a float value defaulting to 5) dictates how strictly Ovis Image follows the text prompt. You can adjust this between 3.0 and 10.0, where higher values force Ovis Image to adhere closely to the text, while lower values allow for more creative interpretation.

Resolution and Configuration

To control the visual dimensions, Ovis Image uses the image_size parameter. This allows you to select from various enum options including square_hd, square, portrait_4_3, portrait_16_9, landscape_4_3, and landscape_16_9. By default, Ovis Image uses landscape_4_3.

For reproducibility, Ovis Image accepts a seed integer (from 0 to MAX). Setting a specific seed allows you to reproduce the exact same image in future runs. Finally, you can determine the file type using the output_format parameter. Ovis Image defaults to png, but also supports jpeg and webp formats.

Recommended settings

Photorealistic Photography

Use 28–35 steps and a Guidance Scale of 5–6. Ensure your prompt includes camera details (e.g., "iPhone photograph," "natural lighting") and specific textures to get the best results from Ovis Image.

Complex Scenes

If your prompt involves multiple subjects or specific spatial arrangements, increase the Guidance Scale to 7.0 to force Ovis Image to strictly adhere to the text description.

Speed Optimization

For rapid iteration with Ovis Image, reduce num_inference_steps to 20 and use jpeg output format to minimize latency and file size.

Output quality and performance

Ovis Image targets high-fidelity outputs suitable for commercial and creative use.

On RunComfy, expect:

• Visual Fidelity: Sharp focus, realistic skin tones, and accurate material rendering (fabric, nature, metal).
• Latency: Generation typically completes in seconds, optimized by RunComfy's high-availability GPU clusters.
• Consistency: High reliability in reproducing styles across different seeds when the Ovis Image prompt remains constant.

For best stability:

• Use the provided image_size presets (e.g., landscape_4_3) rather than custom pixel dimensions to ensure Ovis Image stays within its training distribution.
• Utilize the negative_prompt to scrub generic digital artifacts.

Recommended use cases

Ovis Image excels at scenarios requiring high semantic understanding:

Lifestyle & Fashion

Generate "gorpcore" or streetwear imagery with specific clothing textures and outdoor backgrounds using Ovis Image.

Digital Marketing

Create unique assets for social media campaigns that require specific brand colors or moods described in text.

Storyboarding

Rapidly visualize scripts or concepts where specific actions and interactions are described in the prompt.

How Ovis Image compares to other models

Ovis Image vs SDXL / Flux

• Ovis Image: Often demonstrates superior understanding of complex sentence structures due to AIDC-AI's multimodal training techniques.
• Flux: Known for extreme typography generation; Ovis Image may offer a different aesthetic focused on natural scene composition.

Ovis Image on RunComfy vs Local Setups

• Running Ovis Image locally requires significant VRAM and environment configuration.
• RunComfy provides instant access via Playground and API, handling all dependencies and GPU scaling automatically.

Official resources and licensing

Official AIDC-AI Hugging Face

https://huggingface.co/AIDC-AI

Official GitHub

https://github.com/AIDC-AI/Ovis

License and commercial usage

Ovis Image models generally follow the licensing terms provided by AIDC-AI. Users should verify the specific model license on the official Hugging Face repository before engaging in large-scale commercial applications.

RunComfy facilitates the infrastructure to run these models but does not supersede the original Ovis Image licensing terms.