FLUX.1 Krea image generation: mechanisms and practical controls

Create a new still image from a text description; no uploaded image is required. Begin with the core Positive prompt control and the supplied square, single-image setup. The saved prompt describes a centered portrait with cool lighting and a blurred background, but does not restrict what subjects you can request.

The named result is Sweet Tea Preview Image, port output:9. VAEDecode node 201 supplies IMAGE slot 0 to SweetTeaPreviewImage node 9 through its images input. The interface describes node 9 as no-permanent-save output transport, so a preview is not evidence of a permanently saved file. The pack contains no executed example or comparable run timing.

For an environment running this recipe, match the actual loader selections:

The retained model URL for ae.safetensors points to Comfy-Org's Lumina_Image_2.0_Repackaged repository. Preparation also lists the required tea-time-labs/sweet-tea-nodes node pack.

The VAE resource's transitive preparation dependency is all_in_one/lumina_2.safetensors, provider version 2674864, checksum b703b020a8ff07994e06b3dfa1b28625192bdbfa39aa086a33ea6d6fd206b186. It is not a second diffusion model selected by this graph and must not replace the Krea file in node 203. The pack does not explain how that transitive dependency supplies the selected VAE file. The four directly selected model files have provider version main and no retained artifact checksums.

Model and encoder selection are fixed in the tool interface. Native loader schemas expose other choices: 203.weight_dtype lists default, fp8_e4m3fn, fp8_e4m3fn_fast and fp8_e5m2; 205.device lists default and cpu. Those choices do not establish suitable substitutions or measured benefits for this recipe. Keep the selected setup unless you have separate compatibility evidence. The diffusion filename alone does not establish the precision of the whole workload.

No scoped GPU capacity, system RAM recommendation, supported backend/platform list or comparable duration is retained. The preparation workload records 1024 × 1024 but leaves batch and precision unspecified; the graph separately saves batch_size=1.

The graph connects three stages: text conditioning, latent sampling and image decoding.

Conditioning: 205 supplies CLIP slot 0 to 204.clip. Node 204 encodes text and sends CONDITIONING slot 0 directly to 202.positive, and also to 206.conditioning. Node 206 ConditioningZeroOut supplies CONDITIONING slot 0 to 202.negative. The retained resource describes this as zeroed negative conditioning: there is no independently editable negative-prompt text field.

Sampling: 203 supplies MODEL slot 0 to 202.model. 207 EmptySD3LatentImage creates LATENT slot 0 using width, height and batch_size, then feeds 202.latent_image. KSampler 202 denoises that latent using its conditioning and sampling settings. The interface port named image_count controls 207.batch_size; despite its retained port-role label source_image, it is an integer count, not an uploaded-image input.

Decoding: 202 sends LATENT slot 0 to 201.samples; 208 supplies VAE slot 0 to 201.vae. Node 201 VAEDecode produces IMAGE slot 0, which reaches the preview through 9.images. There is one exposed image output, not separate alternative outputs for the intermediate stages.

The initial latent is empty, not an encoded photograph. Lowering denoise therefore cannot preserve an uploaded image. CFG is classifier-free guidance against the zeroed negative branch, not a separate FLUX guidance setting. The graph has neither an image-input path nor a separate FLUX guidance control. These retained connections explain the processing path; they do not characterize all runtime side effects.

The core interface exposes Positive prompt, Width, Height and Image count. Sampling controls are advanced. These bindings are real scalar inputs in the retained graph; none of the adaptations below replaces a graph connection.

Schema ranges are accepted input ranges, not recommendations for hardware capacity or useful image quality. For aspect-ratio changes, set both dimensions deliberately. A smaller width with the baseline height gives a portrait format; that changes generation dimensions rather than cropping an existing image.

Seed requires a UI/native distinction. The UI default is -1, documented as randomization. The native input instead requires a nonnegative integer and saves ; its schema reports an upper bound of . A retained widget also says . The pack does not establish how the UI translates its sentinel, and the interface's own minimum constraint is 0. Do not write directly to the native input. Use a fixed nonnegative seed to reuse starting noise during comparisons; this is not a guarantee of pixel-identical output across runtimes.

First distinguish a changed setting from a feature this graph does not provide. In particular, node 201 is ordinary VAEDecode, not VAEDecodeTiled; it has connected samples and vae inputs, but no tile_size input. The retained tiled-decode implementation concerns a different node and changes only the decode stage, not model loading or denoising. Applying that technique here would require a graph change, not a scalar-control adjustment. No executable fallback recipe or measured memory remedy is retained.