Skip to main content

Documentation Index

Fetch the complete documentation index at: https://docs.melius.com/llms.txt

Use this file to discover all available pages before exploring further.

This page collects the issues we see most across marketer onboarding sessions, and the fixes that work. If something on this page describes what you’re hitting, the fix should take less than a minute. If it doesn’t, drop a note in the Slack channel and we’ll help directly.

Output doesn’t reflect the style anchor I connected

Most common cause: you didn’t run the text node. The Image Style Analysis template, and any text node with an LLM call, doesn’t auto-run when you create it. You have to click into the node and hit Run. If you connect an image node to a style analysis that’s never been run, the image node sees an empty context and falls back to whatever its own prompt says. Fix: click the style analysis text node, hit Run, wait for the output to appear, then re-run the downstream image node. Subsequent re-runs will use the same analysis automatically.

Logo is wrong, distorted, or invented

Most common cause: the model is generating the logo from scratch instead of using your file as a reference. Fix:
  • Drag your logo file (PNG or SVG) onto the canvas as its own node.
  • Connect it to the image node as a reference.
  • In the prompt: “Place the @logo as shown in the reference. Do not redraw it — use the file exactly.”
If the logo still drifts, switch to GPT Image 2 (High) — it’s more faithful to reference images for graphic elements.

Product dimensions / proportions look wrong

Most common cause: the model is guessing at your product’s size and shape from a single lifestyle photo. Fix:
  • Paste the PDP URL into the agent chat: “Pull product details from [URL] focusing on dimensions and proportions.” The agent creates a text node with the actual specs.
  • Add a clean pack shot (studio photo on neutral background) as a separate reference.
  • Connect both to the image node.

Output looks blurry on the canvas, but downloads sharp

This is expected behavior. The on-canvas preview is downscaled to keep the canvas fast and responsive. The actual file is full-resolution. To see the real output:
  • Click the image to open it in full-screen mode, OR
  • Download the file — it’ll be at whatever resolution you set on the node (1K / 2K / 4K).
If the downloaded file is also blurry, check the node’s resolution setting — you may be generating at 1K when you wanted 2K.

”I love this generation. The next one isn’t as good.”

Most common cause: image models are probabilistic. The same prompt run twice can produce noticeably different results. Fix:
  • Set the variations count to 3–4 on the image node. Multiple variations run in parallel and give you options.
  • Tighten the prompt with specific constraints. “Centered composition, soft golden-hour light from camera-left, shallow depth of field, terracotta and cream palette” gives the model fewer degrees of freedom than “warm aesthetic.”
  • Use the agent to write the prompt: “Take this image (the one I like) and write a detailed prompt that would reliably reproduce it.” The agent will generate a much more specific prompt than you’d write yourself, which reduces variance.

Agent generated something completely off-brief

Most common causes: vague brief, missing context, or the agent got bored of asking clarifying questions and made assumptions. Fix:
  • Switch the agent from auto-run to ask permission mode. You’ll get a chance to inspect and correct the plan before generation kicks off.
  • In the brief, name the constraints explicitly. Not “make some ads”“5 static ads, 4x5 aspect ratio, hero product shots in the brand’s editorial style, on warm neutral backgrounds, no on-image copy.”
  • Include reference images. The agent’s interpretation of “warm aesthetic” is much sharper when it has two or three example images to anchor against.

Forgetting to run a node before connecting downstream

A recurring issue with text nodes, especially the Image Style Analysis template. The fix is the same as the first item on this page: text nodes need to be Run before their output is usable. Get in the habit of running every text node immediately after you connect its inputs.

I want to change one thing in an image without regenerating the whole thing

You like the image. You want to remove the text, change one element, swap a color, or clear a region. You don’t want to lose the rest. Two approaches; try the first one first. Approach 1: Image-to-image regeneration with a “keep” prompt. Create a new image-to-image node, drag your existing image in as the reference, and prompt:
Recreate this exactly, but [the specific change].
Keep everything else identical.
Concrete examples that work: “Recreate this exactly, but remove all text apart from the @logo.” “Recreate this exactly, but change the background to a peach gradient.” “Recreate this exactly, but make the hero product orientation horizontal.” This is usually the cleanest path. The model preserves what you don’t mention and re-renders the rest. Approach 2: Inpaint. On the image node, open the Inpaint tool. Draw a mask over the area you want to change. In the prompt field, describe what should fill the masked region. Inpaint is improving but isn’t always the best tool. If the area you’re changing is bordered by complex detail, the seam can show. For “remove all text” or “change the background,” the image-to-image route is usually better. Reach for Inpaint when the change is highly localized and the surrounding image needs to stay byte-identical.

”I have so many canvases and I can’t find anything”

Most common cause: no naming convention, no project structure. Fix:
  • One Melius project per brand / client. Don’t mix brands.
  • One canvas per campaign or workflow. Name it after the campaign — “Spring drop — email banners,” not “Untitled.”
  • Archive old canvases into an “Archive” sub-section, or rename them with an “Archive —” prefix.

Mac feels sluggish on a busy canvas

Most common cause: too many image nodes at full resolution preview on a single canvas. Fix:
  • Generate exploratory work at 1K resolution. The on-canvas preview is much lighter.
  • Split very large workflows across multiple canvases in the same project (one canvas per workflow stage, e.g. “Character sheets,” “Clip generation,” “Stitching”).
  • The canvas can comfortably hold 50–100 image nodes. Past that, performance starts to dip. Move finished work into a separate “Completed” canvas if your active canvas gets crowded.

My agent ran out of context / lost track of the canvas

Most common cause: the agent chat has accumulated a lot of back-and-forth in a single thread. Fix:
  • Start a new chat on the canvas for a fresh task. The agent reads the canvas state regardless — it doesn’t need chat history to know what’s on the canvas.
  • You can also tag specific nodes with @ in a new chat to point the agent at exactly what you want to work on. “In a new chat: take @hero-shot and generate 5 lifestyle variants.”

My MCP setup with Claude isn’t working

Most common causes:
  1. API key generated from the wrong team. Re-generate from the team you actually want Claude to access.
  2. Claude doesn’t see the Melius connector. Restart Claude Desktop after adding the connector.
  3. Claude returns auth errors. Re-authorize the connector — the API key may not have been saved on the first try.
If none of these fix it, drop the issue in the Slack channel with a screenshot of the error.

Native AI / EPS / PSD file support

We don’t currently support .ai or .eps for vector files, or .psd for layered editing. Vectors need to be SVG or PNG. We’ve heard the request and are tracking it — likely on the roadmap, but no firm date yet. In the meantime: export your .ai files to SVG (vector) or PNG (raster) for use as references.

I want a feature that doesn’t exist yet

Drop it in the Slack channel. Most feature requests from active customers get evaluated within the week, and many ship within days. We’ve shipped Magic Resize, smart aspect ratios, in-painting, and custom font upload from customer asks. If you’ve got an idea, we want to hear it.
This page gets updated as new gotchas surface. If you hit something not on here that took you more than a few minutes to figure out, ping us in Slack — we’ll add it.
Last modified on May 18, 2026