The workflow is not primarily an image model; it is a prompt-assistant layer inside ComfyUI. A local Ollama model reads your reference image, describes it, and rewrites the description into a concise prompt for a separate image-generation workflow.
InputReference image
Load a photo or generated reference image into the Ollama chat node. The model acts as a local vision/text analyst.
MiddleGemma 4 via Ollama
Gemma 4 produces a cleaned, image-model-friendly prompt rather than a raw long caption.
OutputZ-Image / Flux-style generation
The generated prompt is connected to your positive prompt input, so the image model receives a structured prompt automatically.
Best use: reference-image prompting, AI-influencer character ideation, style transfer prompts, outfit/background variation prompts, and faster prompt iteration in local ComfyUI workflows.
2
Safety and use boundaries
Important: The source video discusses an uncensored/adult-capable model. This public guide preserves the technical setup while avoiding explicit prompt recipes. Use adult-capable models only for lawful, private, fictional, consenting-adult creative experimentation. Do not generate minors, real people without consent, non-consensual scenarios, harassment, impersonation, extortion material, or illegal/exploitative content.
- Use fictional adult subjects only when adult capability is relevant.
- Do not use coworker, student, celebrity, private-person, or family likenesses without explicit permission.
- Do not upload sensitive source images to public services.
- Review outputs before sharing and discard unsafe or unintended results.
- Clearly label AI-generated media when shared.
- Read the model page, license, and prohibited-use terms before downloading or using any uncensored/adult-capable model.
Required
- ComfyUI installed and able to launch.
- ComfyUI Manager installed, or another method for installing custom nodes.
- Ollama installed and running in the background.
- An Ollama model that supports the behavior you want. For image analysis, choose a vision-capable model; confirm on the model page.
- A working image-generation workflow such as Z-Image Turbo, Flux, or another workflow that accepts a positive prompt string.
Assumptions
- The video uses Windows/PowerShell language, but the same Ollama concept works on macOS and Linux with the relevant terminal.
- The exact model/tag shown in a video may change. Use the run command from the current Ollama model page rather than guessing.
- If your image workflow requires specific checkpoints, CLIP/text encoders, or VAEs, install those separately before wiring the prompt assistant.
4
Setup steps from the video
0:42 1. Update ComfyUI first
Open ComfyUI, launch Manager, and update ComfyUI/custom nodes before installing new nodes. This reduces missing-node errors after loading the workflow.
0:52 2. Install the ComfyUI Ollama custom node
- Open Manager → Custom Nodes Manager.
- Search for
ComfyUI Ollama. - Install the node package.
- Restart ComfyUI after installation.
If multiple Ollama node packs appear, choose the one that provides an Ollama chat node and an Ollama connectivity/server node, or use the repository linked in Sources.
1:03 3. Install and start Ollama
- Install Ollama from the official download page.
- Launch the Ollama app/service.
- Confirm it is running before you try to pull or run a model.
# Optional terminal check
ollama --version
ollama list
1:16 4. Pull the model from its Ollama page
The video says to copy the run command from the Ollama model page and paste it into PowerShell/Terminal while Ollama is running. Because model names and tags change, use the current page's exact command.
# Pattern only — replace with the exact command from the model page
ollama run owner/model-name:tag
During source checking, current example pages included Gemma 4 and uncensored/adult-capable variants such as baytout3/Gemma-4-Uncensored-HauhauCS-Aggressive:e4b. Treat these as examples to verify, not universal recommendations.
2:03 5. Return to ComfyUI and add Ollama nodes
- Add an Ollama Chat node.
- Drag from its connectivity/server input and create/load the Ollama Connectivity node.
- Confirm the node points to your local Ollama server, usually
http://127.0.0.1:11434. - Select the downloaded Gemma 4/Ollama model in the model field.
2:24 6. Set keep-alive to protect VRAM
The video recommends setting keep_alive to 0 minutes if VRAM is tight. This unloads the language model after it generates text, leaving more VRAM for your image model.
5
Build the image-to-prompt workflow
- Add a Preview Text node. Connect the Ollama chat output to Preview Text first so you can see exactly what the language model is producing.
- Feed the reference image into the Ollama chat node. The video uses Gemma 4 as a vision-capable analyst. If your selected model cannot accept images, switch to a vision-capable Ollama model or a node pack that supports multimodal input.
- Add a system prompt. The raw image description is often too long for image models. The system prompt should compress the caption into a clean positive prompt.
- Load your image-generation template. The video uses a basic Z-Image Turbo template; a Flux-style workflow can also work if it accepts a positive prompt string.
- Connect text output to positive prompt. Route the Ollama chat text output into the generation workflow's positive prompt input.
- Leave a user prompt box available. Use it to request specific changes such as hairstyle, shot type, outfit, or background while preserving the core reference.
Success check: When you run the workflow, the Preview Text node should show a concise prompt, and the image-generation node should create an image based on that prompt.
6
Safe system-prompt pattern
The exact Patreon system prompt from the creator is not public in the video. Use a safe, non-explicit version like this as a starting point:
You are an image-prompt assistant for ComfyUI. Analyze the provided reference image and write one concise positive prompt for an image-generation model. Preserve the important visual identity, composition, camera angle, lighting, outfit, pose, environment, and style. Keep the output as a comma-separated prompt phrase, not a paragraph. Do not include policy discussion, analysis labels, or unsafe content. If the user asks for a modification, preserve the reference image's core identity and apply only that requested change.
Then use the user prompt field for small creative direction:
Preserve the character, but change the hair to long dark hair.Make this read as a mirror selfie while preserving the outfit and room.Turn this into a regular selfie with the same lighting and casual style.Keep the character consistent, but change the background to a bright studio.
Do not use this guide to create explicit prompt recipes. If you experiment with adult-capable models, keep it lawful, fictional, consenting-adult, private, and compliant with model/platform terms.
- Ollama check:
ollama list shows your model, and the Ollama app/service is running. - ComfyUI node check: Ollama Chat and Ollama Connectivity nodes appear after restart.
- Connection check: The connectivity node points to the local Ollama server.
- Text check: Preview Text displays a short, clean image prompt.
- Generation check: The Z-Image/Flux workflow produces an image without missing-node, missing-model, or VRAM errors.
- Reference check: The output preserves the intended composition/style, but do not expect perfect identity without character LoRAs or identity-control tools.
- Ollama model is not listed in ComfyUI: make sure Ollama is running, the model was pulled successfully, and ComfyUI was restarted after installing the custom node.
- Connection refused or timeout: verify the Ollama server URL, usually
http://127.0.0.1:11434, and confirm no firewall/VPN rule blocks local connections. - Missing custom nodes: update ComfyUI, reinstall the Ollama node pack, restart ComfyUI, and reload the workflow.
- VRAM spikes or image model fails after prompt generation: set
keep_alive to 0, reduce image resolution, use a smaller quantized Ollama model, or run the LLM on CPU/system RAM if supported. - Generated prompt is too long: tighten the system prompt: “Output one concise comma-separated positive prompt under 90 words.”
- Image does not preserve the same face/character: this workflow improves prompt transfer, not identity lock. Use character LoRAs, reference adapters, IP-Adapter-style nodes, or identity-control workflows if exact consistency matters.
- Mirror selfie or shot type is missed: put the shot type in the user prompt field, e.g.,
make this a mirror selfie. - Unsafe or unwanted content appears: add explicit negative prompts and safety instructions, review outputs before sharing, and switch to a safer model if needed.
9
Sources and reference links
Source notes: transcript fetched locally from the YouTube URL on 2026-06-07; description links and practical steps were preserved. Model examples were checked against current Ollama search/model pages, but readers should verify the exact model/tag/license before use.