Local image generation can support a business preparing visual concepts, campaign variants or catalogue materials. The computer executes the model while the team controls references, settings and approval. Organise implementation as a production workflow from brief to accepted file, checking each model variant's commercial conditions.
Choose the material you need first
A campaign concept, product composition and explanatory image have different acceptance criteria. Concepts need a suitable visual direction and adherence to the brief. Product work must preserve actual shape, colour and required elements. Informational material must convey the idea accurately and remain legible at its display dimensions.
Prepare a brief defining purpose, audience, format, visual identity and mandatory elements. Select reference images the business has permission to use. When an image illustrates an actual product or service, compare the result with what the business can deliver. Keep demonstration scenes distinct from photographs documenting real projects and results.
The business visual workflow
- Brief
Purpose, format, references and required elements.
- Local workflow
Identified components and parameters.
- Review
Fidelity, product, text and consistency.
- Approval
Final file and configuration record.
Select the model and exact variant's licence
FLUX.2 klein 4B is one example supporting local generation and editing. Black Forest Labs publishes 4B and 4B Base under Apache 2.0. The same page lists 9B and 9B Base under a Non-Commercial licence. Model-family selection must continue to the actual variant and artefact licence.
Record the main model, text encoder, VAE and any adapters separately. Retain notices and conditions for installed components. A permissive main-model licence does not replace checking input photographs, fonts or materials subsequently added to a composition. Keep that record with the visual project.
Configure a verifiable local workflow
ComfyUI documents text-to-image and editing workflows for FLUX.2 klein. In the 4B example, files occupy separate directories: the encoder in models/text_encoders, diffusion model in models/diffusion_models and VAE in models/vae. The workflow must reference installed files and an application version containing the necessary nodes.
A business assessment profile can use 64 GB RAM, a 2 TB SSD and a dedicated 16 GB GPU. Actual memory depends on components, precision, resolution and active operations. Start with one request, the agreed pilot resolution and an official workflow. Measure memory consumption and completion time before adding batches or additional nodes.
Distilled and Base serve different roles
Distilled variants target generation with fewer steps, while Base variants provide a foundation for other configurations and adaptation. Steps, sampler and settings must match the variant. Copying another model's workflow can create an unsuitable combination even when the application loads every file.
Compare two configurations with the same brief and references. Keep parameters and output images, then check mandatory elements. Supplier timings for a particular GPU belong to those stated conditions. Measure time and memory separately in acceptance trials on the business computer.
Check the model variant
- 4B / 4B Base
Apache 2.0 on the reviewed supplier page.
- 9B / 9B Base
Separately published Non-Commercial conditions.
- Components
Keep separate encoder, VAE and adapter records.
Build editing around the correct photograph
For editing, retain the original and describe the required change precisely: background, composition or a demonstration element. For real objects, compare geometry, branding, packaging and characteristics after each variant. Review an attractive image that changes the product before using it commercially.
Check text, prices and specifications against approved data and add them with suitable design tools where appropriate. Text drawn into an image is not an exact commercial record. Keep composition versions and source files so pricing or identity corrections can be made in a controlled way.
Manage batches and variants
Each job receives a brief, references, model, workflow, seed and relevant parameters. The seed is one part of the configuration; comparisons also retain software versions, components and backend. Name variants so the team can identify approved work and experimental outputs.
For a small team, a queue with one active job can simplify sharing a GPU between image production and other local models. Define who may start batches, their size and retained SSD space. After selection, archive useful files and apply the working-variant retention policy.
What to retain for each job
- Input
Brief and authorised references.
- Execution
Models, workflow, seed, versions and parameters.
- Output
Variants, corrections and the approved file.
Acceptance includes the final material
Assessment covers brief fidelity, product characteristics, legibility and delivery dimensions. Inspect material at the size used on the website or campaign. For a shop, compare consistency across images of the same product. Prepare the final file for its channel and keep the approved version with its production record.
Handing a workflow over to the team
Save the workflow alongside its required model files and nodes. Document directories, versions and storage use so another authorised person can reconstruct the environment. Additional nodes are software components: review their source, dependencies and connections before adding them to business work.
During acceptance, restart the application, load a demonstration reference and execute the retained trial job. Check the output file and recovery from saved configuration. Next test the operator account with the permissions actually intended for use. Keep references, working variants and approved images separate. This organisation lets the team continue production and correct a job without reconstructing settings from memory. Include the handover trial in future upgrade checks, retaining any relevant differences before approving the changed workflow.
WebHat uses a customised local AI system and provides consultancy on equipment selection and adapting visual workflows to business work. A custom application with AI integration can organise briefs, queues and review. To connect a catalogue or material-management system, design business automation around existing data and approvals.
Questions about locally generated images
Do all variants in a model family have the same licence?
Check each variant. In the FLUX.2 klein example, 4B is published under Apache 2.0 while 9B has separate Non-Commercial conditions.
Is a seed enough to document an output?
Retain model, workflow, settings, references and versions as well. Technical comparisons use the relevant complete configuration rather than the seed alone.
Can a product photograph be a reference?
Yes, where the business has permission to use it. Compare the result with the actual product and verify commercial elements and characteristics before approval.
How can equipment be shared with other AI applications?
Schedule jobs and check memory. A queue with explicit limits supports sequential execution on the shared GPU.
