Qwen Image 2512 text to image: mechanisms and practical controls

Begin with the supplied standard configuration: 1328 × 1328 pixels, image count 1, Lightning mode off, 50 standard steps and standard CFG 4. The shared sampling settings are Euler, simple, sampling shift 3.1 and denoise 1. No uploaded image is required. Review the populated negative prompt for features you actually want to discourage.

The named result is Generated images, port output:60. Decoded IMAGE data from node 244, slot 0, enters the images input of node 60 (SweetTeaPreviewImage), whose terminal metadata specifies a batch. The retained interface describes the terminal as image preview transport without permanent saving. It does not establish an exported file format, a download procedure or a complete characterization of runtime side effects.

The example prompt's request for “32K” precision is text conditioning. Width and height set the requested dimensions; the graph contains no upscaling stage.

The recipe fixes these loader selections; the current creator interface exposes no model-selection controls.

Node and class	Actual inputs	Directory recorded in the graph
248 — UNETLoader	unet_name: qwen_image_2512_fp8_e4m3fn.safetensors; weight_dtype: default	diffusion_models
252 — CLIPLoader	clip_name: qwen_2.5_vl_7b_fp8_scaled.safetensors; type: qwen_image; device: default	text_encoders
261 — VAELoader	vae_name: qwen_image_vae.safetensors	vae
249 — LoraLoaderModelOnly	lora_name: Qwen-Image-2512-Lightning-4steps-V1.0-fp32.safetensors; strength_model: 1	loras

Preparation also lists tea-time-labs/sweet-tea-nodes as a required node pack; the graph needs SweetTeaPreviewImage for its output. The runtime must also provide the retained core node classes, including ComfySwitchNode and the primitive nodes used to route settings.

All five preparation components are recorded as required, including the Lightning LoRA. Standard mode selects the base-model branch, but the retained evidence does not establish whether runtime validation or execution allows the inactive LoRA artifact to be absent. Keep that distinction when diagnosing a missing-file error.

The diffusion model and text encoder have unresolved Resource identities. The LoRA, VAE and node pack have resolved Resource identities, but no retained component checksums or node-pack commit pin. Model provider versions are recorded as main: exact filenames are known, immutable artifact versions are not. Loader inventories are supplied by the runtime. Neither the filenames nor weight_dtype: default establish the precision of every execution stage.

No scoped GPU capacity, system RAM recommendation, comparable duration, backend list or platform list is retained. The preparation record labels a 1024 × 1024 workload, while the graph and interface default to 1328 × 1328. That preparation label is not a measured capacity profile for either size. Node metadata records several comfy-core versions but does not establish a minimum compatible installation version.

Node 252 supplies the text encoder to both CLIPTextEncode nodes. Node 251 encodes positive text; node 253 encodes negative text. Their CONDITIONING outputs feed node 245 (KSampler) through positive and negative. The supplied Chinese negative text discourages low quality, anatomical distortions, oversaturation, waxy skin, confused composition and distorted text. These are conditioning requests rather than guaranteed corrections.

Node 254 (EmptySD3LatentImage) supplies the starting LATENT using width, height and batch_size. Its output feeds KSampler.latent_image. Although the retained port contract labels batch_size with role source_image, its modality is integer and its actual binding is the image-count input. There is no uploaded-image port, source-image loader or image-to-latent encoding path. Lowering denoise therefore cannot preserve an uploaded image or enable editing. The SD3 and CLIP class names also do not change the fixed Qwen model and encoder selections.

Node 248 supplies the base MODEL. The model selected by node 247 passes through node 246 (ModelSamplingAuraFlow) and then into KSampler.model. The interface identifies node 246's shift as a sampling schedule shift. It provides no directional prediction for the effect of increasing or decreasing that value.

KSampler uses the selected model, conditioning, noise seed, steps, CFG, sampler, scheduler and denoise to produce LATENT samples. Node 244 (VAEDecode) receives those samples and the VAE from node 261 and produces IMAGE data for the terminal. The actual decoder is VAEDecode; this selected recipe contains no VAEDecodeTiled node, tile-size input or overlap input.

The graph supports describing this task as generate_image. The retained operation record has no compiled capabilities, rules or adaptations, so that record does not add guarantees about editing, identity preservation or original-pixel preservation.

The lightning_mode port binds to node 250 (PrimitiveBoolean), input value, default false. Its output controls three ComfySwitchNode nodes: 247 selects the model, 255 selects steps and 258 selects CFG.

Mode	Model selected by 247	Steps selected by 255	CFG selected by 258
false: standard	Base MODEL from 248	Node 256 value: 50	Node 259 value: 4
true: Lightning	LoRA-modified MODEL from 249	Node 257 value: 4	Node 260 value: 1

Only the selected branch's steps and CFG reach KSampler. Both branches retain their own settings, so switching modes does not reset values you previously changed. Editing standard steps while Lightning is enabled has no sampling effect, and editing Lightning CFG while standard mode is selected has no sampling effect.

The lightning_strength port binds to node 249 (LoraLoaderModelOnly), input strength_model, default 1, range −100 to 100 with step 0.01. It affects the selected model only in Lightning mode. At strength 0, the LoRA's model contribution is removed while Lightning steps and CFG remain selected. This does not restore the standard configuration.

At active CFG 1, negative conditioning does not influence sampling. This includes default Lightning mode and standard mode if its CFG is changed to 1. The negative field remains present and connected; its presence alone does not establish influence.

The supplied Lightning configuration selects four denoising steps instead of fifty and also changes the model and guidance. Changing modes can change appearance. There is no measured speed comparison or retained evidence that another Lightning step count, CFG or strength improves quality.

These controls apply in either mode.

Port	Actual node input	Supplied setting and consequence
seed	245 KSampler — seed	Interface default −1 requests fresh randomness; graph value is 464857551335368. The seed determines the noise used for sampling.
sampler	245 KSampler — sampler_name	euler; the sampling algorithm can change appearance. Choices come from runtime.
scheduler	245 KSampler — scheduler	simple; controls the noise-removal schedule. Choices come from runtime.
sampling_shift	246 ModelSamplingAuraFlow — shift	Graph value 3.1000000000000005, described as 3.1; changes sampling schedule shift. Range 0–100, step 0.01.
denoise	245 KSampler — denoise	1; range 0–1, step 0.01. Lower values change sampling from the empty latent and do not add image editing.

The interface's −1 seed default conflicts with its displayed minimum of 0 and KSampler's nonnegative seed schema. The interface note documents −1 as a fresh-randomness convention, but the pack does not show its conversion into a runtime seed. Use a nonnegative fixed seed for a direct node-bound comparison. Do not pass −1 directly to KSampler on the assumption that its native schema accepts it. The retained graph's randomize widget is marked non-runtime and is not an additional executable input. A fixed seed helps compare settings but does not establish reproducibility across runtimes.

Mode-specific ports map as follows: standard_steps → node 256 PrimitiveInt.value; lightning_steps → node 257 PrimitiveInt.value; standard_cfg → node 259 PrimitiveFloat.value; lightning_cfg → node 260 PrimitiveFloat.value. Their primitive fields expose broader ranges than the downstream sampler. Effective steps must be integers from 1 to 10000; CFG must be 0–100, with a displayed step of 0.1. Those schema limits are not recommended operating values.

Higher CFG increases prompt pressure, but excessive guidance can harm quality. More steps change the denoising step count; the sources establish no universal quality improvement. Retain Euler, simple, shift 3.1 and denoise 1 as the supplied starting point. No alternative sampler inventory, preferred schedule-shift direction or measured optimum is retained.

The width, height and batch_size ports bind to node 254 (EmptySD3LatentImage) inputs of the same names. Width and height default to 1328 pixels, must be at least 16 and use multiples of 16. Together they choose aspect ratio. Retained examples include 1664 × 928 landscape, 928 × 1664 portrait and 1472 × 1104 for 4:3. These are configuration examples, not executed results or capacity recommendations.

Image count defaults to 1 and has a schema range of 1–4096. The upper limit does not establish that a machine can run such a batch. Larger dimensions and batches require more memory; no quantitative capacity or duration estimate is retained. The images travel through the same decode and Generated images output path.

Reducing dimensions changes the requested image size and can change the result. Reducing image count requests fewer images. Neither intervention establishes that a particular runtime will have enough memory. Distinguish loading, sampling and decode failures before choosing a stage-specific intervention.

The retained tiled-decode source describes a different decoder that changes the decode stage, not model loading or denoising. Using it here would require a graph change beyond the exposed controls. Node 244 is VAEDecode, so setting tile_size on it is not a valid adaptation of this selected revision.