Nodaro DocsImage Critic
Score an image for realism, anatomy, prompt adherence, style or character consistency, and route it by the result to build automatic quality checks.
The Image Critic node scores an image with a vision model and tells you what to fix. You choose what to check, such as realism, anatomy or character consistency, and set a pass mark. The node returns a score from 0 to 1 with short feedback. It sends the feedback out through Approved or Rejected, so the rest of the workflow can act on the verdict.
When to use it
- Check every result of Generate Image and send the failures to a fix step.
- Check that every generated image of a character matches its reference picture.
- Score images against a brand's style reference.
- Catch broken hands, eyes or faces before Lip Sync or Motion Transfer.
- Compare several candidates by their aesthetic score.
Quick start
Add the node
Press Tab on the canvas and choose Image › Understand › Image Critic.
Connect the image and what to compare with
Drag the image to check into the Image input. For Character consistency and Style match, also wire the reference picture into Reference. For Prompt adherence, wire the prompt into Prompt, or type it in the settings panel.
Choose the check
Open the settings panel. Choose a Mode and set the Threshold, the score an image needs to pass.
Connect the outputs and run
Wire Approved to the step for good images and Rejected to the fix step. Connecting only one of them is fine. Click Run.
Inputs
| Input | Accepts | What it does |
|---|---|---|
| Image | Image nodes, such as Generate Image, Modify Image and Upload Image | The image to score. Required. |
| Reference | Image nodes, such as Upload Image | The image to compare with. Needed for Character consistency and Style match, and used by All checks when it is connected. |
| Prompt | Text nodes, such as Text | The prompt the image was meant to follow. Needed for Prompt adherence, and used by All checks when it is set. A wired prompt wins over a typed one. |
Image Critic has two outputs. Both carry the same feedback text, and only one of them carries it on each run:
| Output | Carries the feedback when |
|---|---|
| Approved | The score is equal to or above the threshold |
| Rejected | The score is below the threshold |
Modes
| Mode | What it scores | Needs |
|---|---|---|
| Realism | Photorealism. It calls out plastic skin, broken anatomy and impossible lighting. | Image |
| Anatomy | Typical AI failures: hands, eyes, extra limbs, broken faces | Image |
| Aesthetic | Composition, lighting, framing and color, graded like cinematography | Image |
| Prompt adherence | Whether the image shows what the prompt asked for | Image and prompt |
| Character consistency | Whether two images show the same person: face geometry, hair and defining features | Image and reference |
| Style match | How closely the palette, mood and treatment follow a style reference | Image and reference |
| All checks | Every check that the connected inputs allow. The score is the lowest of the checks. | Image, plus the reference and the prompt when you have them |
Settings
| Setting | What it does |
|---|---|
| Mode | What to check. The default is Realism. See Modes. |
| Threshold | The pass mark, from 0 to 1. The default is 0.7. An image is approved when its score is equal to or above the threshold. |
| Prompt (or wire via input edge) | The prompt to check against. It appears for Prompt adherence and All checks. |
| AI Model | The vision model. The default is Claude Sonnet 5.5. |
| Reasoning Effort | How much the model reasons before it answers, on models that support it. The default is Auto (model default). |
| Advanced mode | Gemini and Claude models. Runs the model on the provider directly, so that Temperature and Max Tokens take effect. The call is then charged at the rate of the maker's own service. |
| Pre & post text | Text that is always added before and after the prompt. It is hidden from people who use your workflow as an app. See Prompt pre and post text. |
Changing the mode after a run clears the old score, so a score never belongs to another check. Changing only the threshold re-checks the existing score at once, without a new run.
Build a fix step
The feedback is written as short instructions, such as "Reshape the left hand", so it works as an edit prompt.
- Set the Threshold near 0.7.
- Wire Rejected into the prompt of the node that fixes the image. For example, use the Prompt input of a second Generate Image node that has the first image as a reference.
Both outputs carry text, not the image. Take the image itself from the node that generated it.
For Prompt adherence, wire the same prompt that you gave the generation node. A different prompt gives a misleading score. For Character consistency, wire the reference straight from the source of truth, such as an Upload Image node or the character's approved picture. A regenerated copy may have drifted.
Credits
A run is one call to the vision model. The call is charged for the tokens it actually used, at the selected model's rate. The tokens are the image or images, your prompt and the critique.
- Before the run, the Run button shows an estimate (
≈and a number) of a typical call on the selected model. Gemini 3.6 Flash costs less than the default, Claude Sonnet 5.5, and Claude Opus 5.5 costs more. - A hold comes first. When the run starts, Nodaro holds enough credits for the longest answer the model may write. After the call, the run is charged what it used.
- Reasoning Effort has no price of its own. A higher level usually writes more tokens, and those tokens are charged.
- Advanced mode, and any reasoning effort on a Claude model, run the call on the maker's own service, at that service's rate.
- Failures. A critique that the model's maker billed but that came back unusable is charged for that usage, and the error says how much. A failure that was not billed is refunded.
The All checks mode scores every check in one call, not one call per check. The full rules are in How LLM calls are charged.
Tips
- Score everything at once. All checks scores every dimension in one call.
- Expect English feedback. The feedback is written in English, whatever the language of your editor.
- Tune the pass mark without paying again. Changing the threshold re-uses the existing score.
- Pick the best of several. To choose one image from many candidates, see Choose Best. For text checks, use QA Check.
Frequently asked questions
What does Image Critic return?
A score from 0 to 1 and one to three short sentences of feedback. The feedback leaves through Approved when the score reaches your threshold, and through Rejected when it does not.
How many credits does Image Critic cost?
A run is one call to the vision model, charged for the tokens it used at the model's rate. The Run button shows an estimate of a typical call before you run. The All checks mode scores every check in one call, not one call per check.
Which mode should I use?
Use Realism or Anatomy to catch typical AI mistakes, Character consistency or Style match to compare with a reference image, Prompt adherence to check the prompt, and All checks to run every check your inputs allow.
Can I check that a character looks like the reference?
Yes. Choose Character consistency and wire the reference picture into the Reference input. The critic checks whether both images show the same person, from the face geometry, the hair and the defining features.
What language is the feedback in?
English, whatever the language of your editor.
Related
Generate Image
Modify Image
QA Check
Choose Best
Consistent characters
Last updated on
Describe to Picker
Analyze a photo with a vision model and fill the Person, Styling, Framing, Lens and Camera / Film Stock pickers you connect with valid catalog choices.
Video nodes
Every video node in Nodaro, from text-to-video and talking avatars to extending, cutting, sound, titles, captions, formatting, analysis and 3D scenes.