WorkflowListedExternal

Reference Fill (flux-fill-dev + deepseek-Janus-Pro)

If you can't run, try in Nordy

Sourcecivitai.com/models/1297490

TaskMultiCapabilitiesImage edit · Image to imageFamilyFLUXEstimated VRAMEstimate unavailableBgrade

Runtime profile

Access: External workflow
Status: Listed
Source: Civitai
Updated: Jul 20, 2026

Source description

If you can't run, try in Nordy

🚀 Ultra-Precise Reference-Based Image Generation !

Struggling to get the perfect results with IP-Adapter or fine-tuned LoRA? Try this workflow!

Now, you can easily generate images with an astonishingly similar feel to your reference —without any additional training or IP-Adapter.

✅ Effortlessly create images that match your reference with extreme accuracy

✅ High-quality results without LoRA or extra fine-tuning

✅ Ideal for restoring and extending reference image details

AI-generated commentary

AI-generated explanation based on source and configuration details. Suggestions are clearly labeled.

Purpose

Reference Fill (flux-fill-dev + deepseek-Janus-Pro) is for generating and editing images from a reference image; it can restore or extend image details with masks and text conditioning, then preview or save the generated results.

Flow

The workflow loads and resizes an image, prepares masks, applies outpainting or inpainting conditioning, guides Flux sampling, and decodes the result.

Inputs and outputs

It takes a reference image and includes mask conversion and mask-based cutting for image edits.
Text prompts can be encoded for conditioning, and results can be previewed or saved.

Setup

The listing identifies Nouvo Ai as the publisher and Flux.1 D as the base model.

Dependencies

Named model files include ae.safetensors, clip_l.safetensors, flux1-fill-dev.safetensors, and t5xxl_fp8_e4m3fn.safetensors; the workflow also names deepseek-ai/Janus-Pro-1B.
Named stages include CLIPTextEncode, Cut By Mask, DifferentialDiffusion, DualCLIPLoader, easy showAnything, FluxGuidance, ImagePadForOutpaint, ImageRemoveAlpha+, ImageResize+, and InpaintModelConditioning.
Other named stages are JanusImageUnderstanding, JanusModelLoader, KSampler, LoadImage, MaskToImage, PreviewImage, SaveImage, UNETLoader, VAEDecode, and VAELoader.

Configuration

The description says reference-based generation can be used without additional training, IP-Adapter, LoRA, or extra fine-tuning.
For restoration or extension work, consider ImagePadForOutpaint, InpaintModelConditioning, DifferentialDiffusion, and ImageResize+.
Suggestion · not verified

Cautions

Before running it, verify that your setup includes deepseek-ai/Janus-Pro-1B, ae.safetensors, CLIPTextEncode, and SaveImage.
Suggestion · not verified
The workflow's license is not stated.

Does this need to be edited?

Sources

1 source

CivitaiPrimary

civitai.com/models/1297490

Verified Jul 20

Source excerpts

2 excerpts

NoteJul 20civitai.com/models/1297490

Source context: 310 downloads · Type Workflows · Base model Flux.1 D

NoteJul 16civitai.com/models/1297490