An AI art moodboard should assign one observable job to every reference: palette, material, silhouette, light, composition, or atmosphere. If two images do the same job, remove one. If a reference cannot be translated into an instruction you can audit in the output, it is inspiration—not control.
This page is a living workflow record, updated September 5, 2026. At 02:10 PDT, Google Autocomplete still expanded “ai moodboard” into generator, free, creator, tool, from-image, and interior-design variants. The query still asks for a usable reference system rather than a pile of attractive pictures.
A Krea 2 ComfyUI workflow advertising moodboard, style transfer, multi-reference editing, ControlNet, detailers, and low-memory paths published v16.1.1 on September 4 at 18:35 PDT. When observed, the version showed 53 downloads and five thumbs-up; the resource showed 14,249 cumulative downloads and 335 thumbs-up. It ranked 24th in Civitai's Most Downloaded / Week response and 72nd in Highest Rated / Week. Compared with the September 3 snapshot recorded here, the parent total had added 362 downloads and 14 thumbs-up across roughly 48 hours—not a 24-hour count.
A second signal was narrower but useful: a Krea 2 attribute-control series for contrast, brightness, and saturation ranked 46th in Most Downloaded / Day and 94th in Highest Rated / Day. Its Contrast version had been published September 4 at 06:13 PDT and showed 225 downloads and 18 thumbs-up. The durable lesson is not the LoRA itself. It is that reference coherence and image attributes are separate review layers.
Krea's official moodboard guide separates moodboards from style references: style references accept up to four images and target style transfer, while moodboards can use larger sets and analyze concepts, expressions, characters, and overall atmosphere. That makes reference count a poor proxy for control. Coherence and assigned jobs matter more.
The first experiment: one radio, three reference jobs
We designed a fictional portable radio for a desert field researcher and held its basic silhouette, location, and product category constant. In the first column, rust, cream, charcoal, and faded orange control palette. In the second, brushed metal, ribbed canvas, matte rubber, and scratched glass control material. In the third, hard sun, low horizon, wide negative space, and blowing dust control light and composition.
The generated plate is not a benchmark of a named model. It is an art-direction experiment made in one editorial session. The radio changes more than the brief requested—especially in dial geometry—which shows why a moodboard cannot replace an identity or structure lock. References influence the visual prior; they do not guarantee invariant product design.
The August 26 subtraction test
The new six-cell plate keeps the radio silhouette, handle, speaker, dial, camera, and desert constant. The sequence adds one reference family at a time: neutral structure, rust-and-cream palette, brushed metal and canvas, conflicting overcast plus hard-noon light, a duplicate warm palette, and the resolved board after subtraction.
Two failures are instructive. Conflicting light lowers contrast without making the scene meaningfully softer, because “overcast” and “hard noon” compete for shadow behavior. The duplicate warm palette turns the entire object and environment orange, destroying the separation between accent, body, and landscape. The final frame does not add a seventh reference. It removes the two that cannot own a unique job.
This comparison is an editorial simulation, not a numerical multi-reference benchmark. It gives you an audit method: add one family, inspect one job, and run a subtraction test before accepting the combined board.
The September 3 coherence test: more is not more controlled
We held a fictional teapot, cup, camera, and studio setup constant. The first cell is neutral. The next three add rust-and-cream palette, matte clay surface, and soft side-window light one at a time. Each addition is easy to audit because it owns one observable job.
The fifth cell is the deliberate failure. Glossy chrome claims the same material slot as matte clay, while saturated magenta replaces the approved warm window light. The result is more striking but belongs to a different product world. The sixth cell removes those conflicts and restores the three-job board.
This is an editorial simulation made with OpenAI image generation, not a Krea 2 benchmark. It tests a portable review method: compare reference families through an invariant subject, then reject any addition that moves an already approved job. A tool may accept dozens of images; your brief should still be able to explain why each family is present.
Run the test with three counts: one reference per job, two references per job, and a deliberately mixed board. If the larger aligned set adds useful range without changing approved decisions, keep it. If the mixed board produces an attractive but off-brief world, subtraction has found the problem.
The September 5 attribute audit: separate reference jobs from image knobs
We held one fictional lamp, rust woven shade, matte sand ceramic, walnut table, cream wall, camera, and side light steady. The upper-left is the approved baseline. The upper-middle flattens contrast until the ceramic body merges with the wall. No reference is wrong; the value separation is.
The upper-right increases contrast until the shade and table shadows lose texture. The lower-left raises exposure until the ceramic highlight and wall approach the same empty white. The lower-middle keeps structure but pushes saturation so far that rust becomes loud orange and the wood stops feeling neutral. The final cell restores a readable midtone range without changing the approved material or palette jobs.
This plate adds a review layer that the reference ledger alone did not make explicit. A moodboard can be coherent about what the object is made from and still fail in how the rendered image distributes light and color. Do not repair that failure by adding another palette or material reference. Diagnose the image knob that moved.
Use a two-pass audit after combining references:
- Reference audit: did palette, material, silhouette, light, composition, and atmosphere keep their assigned jobs?
- Attribute audit: are exposure, contrast, saturation, and sharpness supporting those jobs or overpowering them?
Run one-variable corrections. If the ceramic merges into the wall, raise local value separation rather than changing the whole palette. If the shade loses weave, recover shadow detail rather than adding a fabric reference. If rust becomes orange, lower saturation while preserving its role as the accent.
A simple review template:
Preserve approved references: sand ceramic, rust textile, walnut, soft side light. Change only: midtone contrast between lamp body and wall. Keep: exposure, saturation, camera, geometry, materials, and palette unchanged. Reject: crushed shade texture, clipped ceramic highlight, orange color shift.
The generated comparison is not a Krea 2 or LoRA benchmark. It is an OpenAI image-generation simulation of a portable review method. Small weave and ceramic details vary, so the valid observation is the large attribute change, not pixel-level invariance.
Start with a one-sentence decision
Write what the image must make someone understand or feel. For this project:
A repairable field radio should look trustworthy after years of desert use, not luxurious or futuristic.
That sentence rejects glossy black glass, neon interface lighting, and ornamental sci-fi details before the board exists. It also tells us why scratched aluminum belongs while polished chrome does not.
Use a six-slot reference ledger
| Slot | Question | Evidence to collect | Output check |
|---|---|---|---|
| Palette | Which colors dominate, support, and accent? | 3–5 swatches from one lighting family | Colors appear in named roles, not equal shares |
| Material | What is each surface made from? | close crops of metal, fabric, rubber, glass | Edges, wear, and reflections agree with the material |
| Silhouette | What large shape reads at thumbnail size? | one clear object profile | Main mass survives without texture |
| Light | Where is the key light and how hard is it? | one scene with legible shadow direction | Highlight and shadow direction are consistent |
| Composition | Where is the subject and where is the empty space? | one frame with the intended balance | Crop and horizon match the communication job |
| Atmosphere | What particles, weather, or distance behavior exists? | one restrained environmental reference | Atmosphere supports depth without hiding the subject |
Do not fill all six slots automatically. A character turnaround may need silhouette, material, and palette but no atmosphere. A poster may need composition and light more than surface detail. The ledger is a filter, not a quota.
Translate the board before prompting
For every selected reference, complete this sentence:
Use this reference only for [job], specifically [observable property]; do not copy [subject/content].
Examples:
- Use the mineral photo only for palette: sandstone dominant, oxidized orange accent, charcoal details; do not copy rocks.
- Use the canvas close-up only for strap material: ribbed weave, frayed edge, dusty olive; do not turn the radio body into fabric.
- Use the landscape only for composition: low horizon, subject in the lower middle, broad pale sky; do not add the same mountain.
This phrasing matters even in tools that do not expose per-reference weights. It converts taste into a testable assignment and tells you what to change when an output fails.
The FreeArtGen mood board maker is the direct next step for arranging the selected references. Keep the one-job ledger beside the board; the owned tool organizes the inputs but is not evidence for the method.
A reproducible four-pass workflow
Pass 1: structure without style
Generate a neutral version using only the project sentence, silhouette, camera angle, and required components. Reject it if the handle, dial, speaker, or proportions are wrong. Style references make structural diagnosis harder, not easier.
Pass 2: one reference family at a time
Run palette, material, and light as separate tests. Compare results at the same size. Save the best instruction from each family, not necessarily the prettiest whole image.
Pass 3: combine in a fixed order
Add structure, then palette, material, light, composition, and atmosphere. Save each stage. If an output changes an already approved job, stop and repair that conflict before adding another input.
Pass 4: subtract two references
Remove the most redundant reference and the most contradictory reference in separate runs. If either removal improves clarity, the original board had noise. The subtraction test is mandatory precisely because attractive references can still produce an averaged result.
Use this prompt spine after neutral structure works:
Create [deliverable] for [audience/job]. Preserve: [silhouette, components, camera, identity locks]. Palette reference controls only [dominant/support/accent roles]. Material reference controls only [named surfaces and wear behavior]. Lighting reference controls only [direction, hardness, contrast]. Composition reference controls only [crop, horizon, negative space]. Do not copy reference subjects, logos, text, or recognizable designs.
Where moodboards fail
Too many near-duplicates: five beige interiors do not provide five instructions. Keep the clearest one and extract its actual palette.
Subject leakage: a material reference contains a desirable red chair, and every output grows a chair. Crop to the material or explicitly prohibit the source object.
Contradictory light: a soft overcast palette reference and a hard-noon composition reference fight. Name which image controls color and which controls shadows.
Unlicensed copying: a board full of branded products, celebrity faces, franchise characters, or one living artist's work can turn direction into imitation. Use generic material, historical, public-domain, or original references, and describe properties rather than names.
False consistency: multi-reference output can still move knobs, seams, face shape, costume details, or camera geometry. Use a true edit/reference workflow when invariants matter, and audit them separately from style.
Krea's official technical report describes Krea 2 as an open-weights text-to-image family for creative exploration; community workflows show how creators are building more granular reference paths around it. The model is the current signal, but the durable method is simpler: one reference, one job, one output check. Build the ledger, test each family alone, and delete a conflict before you call the board finished.