HiDream I1 Fast image generator: mechanisms and practical controls

Generate a new image by opening this recipe and replacing its example Prompt. The starting configuration is 1024 × 1024 pixels, batch size 1, 16 steps, CFG 1, LCM sampler, normal scheduler, sampling shift 3 and denoising strength 1. No uploaded media is required.

The named result is Generated images, port output:88. It is one batch output, including when the requested batch contains one image. Node 88 (SweetTeaPreviewImage) receives IMAGE slot 0 from node 86 (VAEDecode). The retained terminal metadata mentions an original SaveImage class, but the selected graph uses SweetTeaPreviewImage; it establishes no filename or save-location control. No executed example, measured quality result or timing is retained.

The preparation record requires all six model files and tea-time-labs/sweet-tea-nodes together. The retained loader selections are:

Node and class	Input	Selected file	Model directory	Retained provider version
93 — UNETLoader	unet_name	hidream_i1_fast_fp8.safetensors	diffusion_models	main
81 — VAELoader	vae_name	ae.safetensors	vae	main
82 — QuadrupleCLIPLoader	clip_name1	clip_l_hidream.safetensors	text_encoders	main
82 — QuadrupleCLIPLoader	clip_name2	clip_g_hidream.safetensors	text_encoders	main
82 — QuadrupleCLIPLoader	clip_name3	t5xxl_fp8_e4m3fn_scaled.safetensors	text_encoders	1038246
82 — QuadrupleCLIPLoader	clip_name4	llama_3.1_8b_instruct_fp8_scaled.safetensors	text_encoders	6e6047b116958689a7b94720f6f06f776fc7264f

Node 93 has weight_dtype set to default. The FP8 model filename does not establish a precision setting for every stage. The loaders are hidden and have no public model-selection ports. Their runtime file selectors do not make model substitution a supported interface adaptation.

The node pack supplies a required dependency; the graph's terminal class is SweetTeaPreviewImage. No node-pack version or commit is retained. Artifact checksums are retained for the diffusion model and all four text encoders, but not for the VAE or node pack; the recorded provider version main alone is not an immutable artifact identity.

Graph metadata records comfy-core 0.3.28 on most core nodes and 0.3.56 on node 93. These are node metadata, not a tested compatibility range or a minimum installation version. The pack establishes no backend, platform, GPU capacity, RAM capacity or duration recommendation. Model-file sizes describe storage artifacts and cannot be used as a GPU-memory requirement. The preparation workload leaves batch and precision unspecified; the selected graph separately establishes batch size 1.

Node 82 (QuadrupleCLIPLoader) supplies CLIP slot 0 to the clip inputs of nodes 85 and 90 (CLIPTextEncode). Their text inputs produce CONDITIONING slot 0, connected respectively to positive and negative on node 94 (KSampler). The shared encoder bundle is explicit in the graph. Each encoder's internal contribution and the way its representations are combined are not explained by the retained sources.

Node 84 (EmptySD3LatentImage) supplies LATENT slot 0 to node 94's latent_image. Width, height and batch size configure this new starting latent. Node 93 (UNETLoader) supplies MODEL slot 0 to node 91's model input. Node 91 (ModelSamplingSD3) adjusts the model's noise sampling schedule with shift, then supplies MODEL slot 0 to node 94's model. The sampler uses the model, conditioning, starting latent and seed to produce LATENT slot 0.

Node 86 (VAEDecode) receives the sampler output through samples and node 81's VAE slot 0 through vae. Its IMAGE slot 0 feeds node 88's images input and the public output:88. Sampling and decoding are separate stages. This graph uses VAEDecode, with no tiled decoder, tile_size input or tile-size port.

This is text-to-image generation, matching task family generate_image. There is no image upload, image encoder, mask, reference-conditioning or compositing path. A prompt asking to preserve a subject or label cannot establish original-pixel preservation. Lowering denoise also cannot supply an existing image that is absent from the graph.

The core ports width, height and batch_size bind to identically named inputs on node 84 (EmptySD3LatentImage). Width and height default to 1024 pixels, have a minimum of 16 and use multiples of 16. Their ratio determines image shape. Batch size defaults to 1 and accepts integers from 1 to 4096. That schema maximum does not establish feasible capacity. Larger dimensions and batches require more memory; no amount or timing change is measured here.

The advanced seed port binds to node 94's seed, which determines starting noise. A fixed seed is useful when comparing prompt or sampling changes while keeping dimensions and batch unchanged. It does not establish identical output across different runtimes.

The interface documents -1 as randomizing each run and retains it as the default, while both its numeric constraint and the sampler schema require a nonnegative value. The graph stores seed 833271177511441; its separate randomize widget value is marked non-runtime and is not an executable node input. The pack does not show how the client handles the interface sentinel. Treat -1 as the documented interface convention, not a direct KSampler value. If it is rejected, choose a concrete nonnegative seed; no failure has been demonstrated in the retained pack.

The width adaptation below is an unexecuted example of a valid control change: with height still 1024, width 768 requests a 3:4 portrait shape. It establishes neither preferred model dimensions nor measured output quality.

The public positive_prompt and negative_prompt ports bind to text on nodes 85 and 90. The supplied positive text describes a red tree; the supplied negative text is bad ugly jpeg artifacts. Both branches are connected, but the native interface states that negative conditioning is unused at CFG 1. Editing only the negative text therefore does not influence baseline sampling.

The advanced cfg port binds to node 94's cfg: default 1, range 0–100 in increments of 0.1. Above 1, it applies negative guidance and changes prompt influence. Excessive guidance can harm results. No preferred higher CFG value is retained for this exact Fast model, and the allowed range is not a quality recommendation. Compare a chosen value above 1 against the baseline using the same seed, prompt, dimensions and batch.

The interface describes conventional negative guidance above 1; it does not establish useful behavior for every allowed value below 1. Remaining at the baseline is a reasonable choice when refining the positive description is sufficient.

The remaining sampling controls are advanced. Their exact bindings and retained defaults are:

Public port	Node and input	Default	Retained constraint or choice source
steps	94 — KSampler.steps	16	Integer, 1–10000
sampler	94 — KSampler.sampler_name	lcm	Runtime choices
scheduler	94 — KSampler.scheduler	normal	Runtime choices
shift	91 — ModelSamplingSD3.shift	3	0–100, step 0.01
denoise	94 — KSampler.denoise	1	0–1, step 0.01

steps sets the denoising step count and changes computation required. The sources do not establish that more steps improve this recipe's quality. sampler selects the sampling algorithm, while scheduler controls the progression of noise levels. They interact with steps. Runtime alternatives are not enumerated here, so select only an option actually offered by the runtime; no alternative combination has a retained quality or speed comparison.

shift adjusts the model's noise sampling schedule before node 94 samples. No alternate optimum, precise schedule formula or directional quality effect is established by the retained material. denoise defaults to full denoising; lower values use only part of the noise schedule. Because node 84 supplies an empty latent, the generic KSampler tooltip about retaining an initial image does not apply to an uploaded image in this recipe.

For a useful configuration comparison, keep seed, prompt, dimensions and batch constant and change one sampling control at a time. The controls below restore individual baseline values; restoring one does not reset the others. The template is a documented starting point, not a measured optimum.