Search FLUX 2 Dev ComfyUI workflow JSON schema 2026 Wan 2.1 LTX Video 2.3 ComfyUI node parameters temporal consistency ZImage Turbo fast image generation ComfyUI workflow 20 results

Designing JSON Schemas for ComfyUI Qwen 3B and Wan 2 Video Nodes

Search Wan 2.1 ComfyUI node parameters seed motion strength frame interpolation LTX Video 2.3 ComfyUI workflow JSON STG spatio-temporal guidance node 13 results

Production JSON Schemas and the FLUX 2 Text Rendering Fix

Here is the technical deep-dive for the production engine. These schemas are designed for the Router → Dispatcher → Remediation loop, with each payload being machine-executable by ComfyUI nodes across your GPU hosts.

Architecture Philosophy The core design principle is semantic isolation: cinematic intent is decomposed into non-conflicting instruction planes before JSON serialization. The Router layer assembles high-level intent into model-specific dialects; the Dispatcher validates node compatibility; the Remediation loop adjusts guidance parameters based on artifact detection.
FLUX 2 Dev — Image Generation Schema FLUX 2 Dev supports native JSON-structured prompting with precise color control and multi-reference capabilities . To prevent text-rendering artifacts, typography instructions must be strictly isolated from visual scene descriptions. When text content bleeds into visual conditioning, the Mistral 3 text encoder can over-activate glyph-rendering pathways, producing garbled or "melted" text . 2.1 Optimal Payload Structure JSON

Copy { "metadata": { "model_host": "flux2-dev-01", "job_id": "img-uuid", "remediation_count": 0, "priority": "high" }, "visual_blueprint": { "scene": { "type": "Cinematic Exterior", "environment": "Neon-drenched cyberpunk alleyway, rain-slicked cobblestones", "time_of_day": "Blue Hour", "atmospheric_conditions": "Light fog, volumetric god-rays from signage" }, "subjects": [ { "id": "subject_01", "description": "Figure in worn leather trench coat, back to camera", "position": "Center foreground, rule of thirds intersection", "color_palette": ["#FF2A6D", "#05D9E8", "#01012B"], "material_properties": "Wet leather reflectivity, subsurface scattering on skin" } ], "lighting": { "key_light": "Neon pink rim light from signage, 45° behind subject", "fill_light": "Cyan ambient bounce from wet ground", "contrast_ratio": "High, 8:1" }, "style": { "rendering_engine": "Photorealistic cinematic", "film_stock": "Kodak Vision3 500T", "quality_modifiers": ["8K UHD", "chromatic aberration subtle", "lens flare organic"] } }, "camera_motion": { "rig_type": "Static Lock", "angle": "Low angle, 15° Dutch tilt", "lens_mm": 35, "f_number": "f/1.4", "focus_distance": "2.1m", "depth_of_field": "Shallow, bokeh hexagonal", "motion_blur": "Optical flow, 1/48s shutter simulation" }, "typography_isolation": { "enabled": true, "text_elements": [ { "content": "NIGHT CITY", "position": "Background signage, out of focus", "font_style": "Bold sans-serif, neon tube", "color": "#FF2A6D", "render_priority": "ambient_only" } ], "text_negative_prompt": "gibberish text, blurry letters, melted typography, watermark" }, "comfyui_node_mapping": { "model_loader": { "node_id": "UNETLoader", "model_path": "flux2-dev-fp8.safetensors", "weight_dtype": "fp8_e4m3fn" }, "text_encoder": { "node_id": "CLIPTextEncode", "encoder_type": "mistral3", "conditioning_strategy": "structured_json" }, "sampler": { "node_id": "KSampler", "sampler_name": "euler", "scheduler": "simple", "steps": 28, "cfg": 3.5, "denoise": 1.0 }, "resolution": { "width": 2048, "height": 1152, "batch_size": 1 } } }

2.2 Cinematic Attribute → ComfyUI Node Mapping Table

Cinematic Attribute ComfyUI Node Parameter Value Range / Notes Lens Choice FLUX2_CameraRig lens_mm 24, 35, 50, 85, 135 Depth of Field FLUX2_CameraRig f-number f/1.2–f/22; maps to optical conditioning Camera Angle FLUX2_CameraRig angle Low/high/Dutch tilt; injected as prefix Color Palette FLUX2_PromptAssembler color_palette Hex array; FLUX 2 parses HTML codes natively Lighting Key FLUX2_PromptAssembler lighting Separated to prevent style bleed Text Rendering Isolated CLIPTextEncode typography_isolation Dedicated negative prompt plane Prompt Adherence KSampler cfg 3.0–4.0 optimal for FLUX 2 2.3 Anti-Artifact Strategy To prevent text-rendering artifacts: Isolate text elements in the typography_isolation object, never in visual_blueprint.scene Use render_priority: "ambient_only" to force text into bokeh/background planes Apply a dedicated text negative prompt to the secondary CLIP encode node Keep CFG ≤ 3.5 to avoid over-saturating glyph pathways

Wan 2.1 & LTX 2.3 — Video / I2V Schema Video models require temporal consistency anchors to prevent flicker and drift. Wan 2.1 uses a spatio-temporal VAE , while LTX 2.3 exposes independent Spatio-Temporal Guidance (STG) and cross-modal sync controls via the Multimodal Guider . 3.1 Wan 2.1 Payload JSON

Copy { "metadata": { "model_host": "wan21-video-01", "job_id": "vid-uuid", "pipeline_type": "i2v", "remediation_count": 0 }, "temporal_consistency_anchors": { "seed": 42, "seed_strategy": "fixed_per_generation", "motion_strength": 0.72, "motion_strength_range": [0.3, 1.0], "frame_interpolation": { "enabled": true, "target_fps": 24, "source_fps": 16, "interpolator": "RIFE_VFI", "interpolation_multiplier": 1.5 }, "spatio_temporal_vae": { "tile_size": 32, "temporal_window": 16 } }, "motion_vectors": { "camera_movement": [ { "type": "dolly_in", "start_frame": 0, "end_frame": 48, "velocity_curve": "ease_in_out", "magnitude": 0.4 }, { "type": "pan_right", "start_frame": 24, "end_frame": 72, "velocity_curve": "linear", "magnitude": 0.25 } ], "subject_motion": [ { "subject_id": "subject_01", "action": "turning_to_face_camera", "start_frame": 32, "end_frame": 64 } ] }, "visual_texture_prompts": { "scene_description": "Cinematic interior, art deco ballroom, golden hour through stained glass", "material_textures": [ { "surface": "Marble floor", "properties": "Polished Carrara, soft reflections, caustics" }, { "surface": "Velvet curtains", "properties": "Deep burgundy, pile texture, light absorption" } ], "lighting_texture": "Warm tungsten practicals, volumetric dust particles, soft shadows" }, "negative_visual_textures": { "artifacts": ["flicker", "sudden lighting changes", "texture popping", "watermark"] }, "comfyui_node_mapping": { "model_loader": { "node_id": "WanVideoModelLoader", "model_path": "wan2.1-i2v-720p-fp8.safetensors" }, "image_conditioning": { "node_id": "WanVideoClipVisionEncode", "start_frame": "input_image.png", "vision_model": "clip_vision_h" }, "sampler": { "node_id": "WanVideoSampler", "sampler_name": "uni_pc", "scheduler": "simple", "steps": 25, "cfg": 6.0, "num_frames": 81 }, "decode": { "node_id": "WanVideoVAEDecode", "tile_size": 32 } } }

3.2 LTX 2.3 Payload JSON

Copy { "metadata": { "model_host": "ltx23-video-02", "job_id": "vid-uuid", "pipeline_type": "t2v", "remediation_count": 0 }, "temporal_consistency_anchors": { "seed": 42, "frame_rate": 30, "num_frames": 121, "resolution": { "width": 1280, "height": 704 } }, "motion_vectors": { "camera_movement": [ { "type": "slow_zoom_in", "easing": "smooth", "description": "Camera gently pushes forward through tree canopy" } ], "environmental_motion": [ { "type": "wind_through_leaves", "intensity": "moderate", "description": "Dappled sunlight shifts across forest floor" } ] }, "visual_texture_prompts": { "scene": "Ancient redwood forest, morning mist, god-rays piercing canopy", "textures": { "bark": "Deep furrowed sequoia bark, mossy patches, damp", "foliage": "Fern undergrowth, dewdrops, subsurface scattering on leaves", "atmosphere": "Volumetric fog, pollen motes, anamorphic light streaks" } }, "multimodal_guider": { "video_guidance": { "cfg": 3.0, "stg": 1.0, "cross_modal_sync": 2.5 }, "audio_guidance": { "cfg": 7.0, "stg": 0.0, "cross_modal_sync": 2.5 }, "optimization": { "video_skip_step": 0, "audio_skip_step": 1 } }, "comfyui_node_mapping": { "text_encoder": { "node_id": "GemmaAPITextEncode", "api_fallback": "local_gemma", "prompt": "motion_prompt_only" }, "guider": { "node_id": "MultimodalGuider", "parameters": "GUIDER_PARAMETERS" }, "sampler": { "node_id": "LTXVTextToVideoSampler", "steps": 40, "strength": 1.0 }, "conditioning_cache": { "save_node": "LTXVSaveConditioning", "load_node": "LTXVLoadConditioning", "reuse_across_remediation": true } } }

3.3 Cinematic Attribute → ComfyUI Node Mapping Table

Cinematic Attribute ComfyUI Node Parameter Value Range / Notes Motion Strength WanVideoSampler cfg 5.0–7.0; lower reduces flicker Frame Interpolation RIFE VFI (post) multiplier 1.5× (16→24fps) or 2× Temporal Smoothing MultimodalGuider cross_modal_sync 2.0–3.5; increases frame coherence Artifact Reduction MultimodalGuider stg 0.5–1.5; Spatio-Temporal Guidance Prompt Adherence MultimodalGuider cfg 2.0–5.0 for video; 3.0 default Motion Description GemmaAPITextEncode prompt Motion-only; scene assumed in I2V Conditioning Reuse LTXVSaveConditioning cache_path Saves seconds per remediation pass 3.4 Motion Vector / Visual Texture Isolation To prevent motion vectors from conflicting with visual texture prompts: Wan 2.1: Use WanVideoSampler with motion prompts in the positive conditioning; keep texture descriptions in the start-frame image embedding via WanVideoClipVisionEncode . The spatio-temporal VAE handles the merge . LTX 2.3: Motion prompts go through GemmaAPITextEncode as pure action descriptions ("camera pans left, water ripples") . Visual textures are either baked into the I2V start image or described in the scene prompt, never mixed with motion verbs. Never describe texture in motion prompts — e.g., say "gentle push forward" not "push forward revealing rough bark texture."

ZImage Turbo — Fast-Track Schema ZImage Turbo is a distilled 6B-parameter S3-DiT model optimized for 8-step generation . The minimal viable payload must enforce the Lumina 2 CLIP type and Qwen 3B encoder to prevent black images and text corruption . 4.1 Minimal Viable Payload JSON

Copy { "metadata": { "model_host": "zimage-turbo-01", "job_id": "fast-uuid", "track": "rapid_prototype", "max_latency_ms": 15000 }, "semantic_payload": { "positive_prompt": "Cinematic portrait, warrior queen, Rembrandt lighting, bronze armor, ember particles, shallow depth of field, 85mm lens", "negative_prompt": "blurry, deformed, extra limbs, oversaturated, watermark, text", "text_rendering": null }, "sampler_config": { "steps": 8, "cfg": 3.5, "sampler_name": "dpmpp_sde", "scheduler": "ddim_uniform", "denoise": 1.0 }, "aura_flow": { "shift": 3.0, "shift_range": [2.0, 5.0], "note": "Higher shift increases composition creativity but may wash out detail [^2^]" }, "comfyui_node_mapping": { "model_loader": { "node_id": "UNETLoader", "model_path": "z_image_turbo_fp8.safetensors", "weight_dtype": "fp8_e4m3fn" }, "clip_config": { "node_id": "CLIPLoader", "type": "lumina2", "encoder_path": "qwen_3_4b.safetensors" }, "sampling": { "node_id": "ModelSamplingAuraFlow", "shift": 3.0 }, "latent": { "node_id": "EmptySD3LatentImage", "width": 1024, "height": 1024 }, "decode": { "node_id": "VAEDecode" } } }

4.2 Cinematic Attribute → ComfyUI Node Mapping Table

Cinematic Attribute ComfyUI Node Parameter Value Range / Notes Speed vs Quality UNETLoader model_path z_image_turbo_fp8 (fast) vs z_image_base_bf16 (quality) Prompt Adherence KSampler cfg 3.0–5.0; 3.5 is the sweet spot Composition Control ModelSamplingAuraFlow shift 3.0 default; 5.0+ for creative composition Text Rendering CLIPLoader type Must be lumina2; Qwen 3B encoder Step Count KSampler steps Exactly 8 for Turbo; 30–50 for Base Sampler Choice KSampler sampler_name dpmpp_sde for detail; euler for speed

Router → Dispatcher → Remediation Integration 5.1 Router Logic The Router selects the schema based on intent_type: Table

Intent Schema Host Selection Criteria cinematic_still FLUX 2 Dev ≥16GB VRAM, FP8 support motion_sequence Wan 2.1 / LTX 2.3 ≥24GB VRAM, temporal VAE available rapid_prototype ZImage Turbo Any available host, <15s SLA 5.2 Dispatcher Validation Before dispatch, validate: Node existence: Check comfyui_node_mapping against host node registry VRAM estimation: FLUX 2 (12GB+), LTX 2.3 13B (24GB+), ZImage (8GB+) CLIP compatibility: ZImage requires lumina2; LTX requires gemma 5.3 Remediation Triggers Table

Artifact Detected Remediation Action Parameter Adjustment Text gibberish (FLUX 2) Isolate typography Increase typography_isolation strictness; reduce CFG 0.5 Flicker (Wan 2.1) Reduce guidance Decrease cfg by 1.0; enable frame interpolation Jittery motion (LTX 2.3) Increase cross-modal sync Raise cross_modal_sync to 3.0; reduce cfg to 2.5 Washed out (ZImage) Reduce AuraFlow shift Lower shift from 7.0 to 3.0; switch to dpmpp_sde Black output (ZImage) Fix CLIP config Force type: lumina2 and FP8 checkpoint 5.4 Conditioning Cache for Remediation For LTX 2.3, always enable LTXVSaveConditioning on the first pass. During remediation, the Dispatcher loads cached conditioning via LTXVLoadConditioning instead of re-encoding, saving seconds per iteration .

Summary Table: Cross-Model Cinematic Mapping Table

Cinematic Intent FLUX 2 Dev Node Wan 2.1 Node LTX 2.3 Node ZImage Turbo Node Lens / Focal Length FLUX2_CameraRig.lens_mm N/A (static or motion curve) N/A N/A (prompt only) Aperture / DoF FLUX2_CameraRig.f-number N/A N/A N/A Camera Motion FLUX2_CameraRig.angle motion_vectors.camera_movement motion_vectors.camera_movement N/A Lighting FLUX2_PromptAssembler.lighting visual_texture_prompts.lighting_texture visual_texture_prompts.textures Prompt prefix Temporal Consistency N/A temporal_consistency_anchors + spatio-temporal VAE MultimodalGuider.cross_modal_sync N/A Motion Strength N/A KSampler.cfg (5–7) MultimodalGuider.cfg (2–5) N/A Frame Interpolation N/A RIFE VFI post-process Native 30–50fps N/A Text Rendering typography_isolation Negative prompt Negative prompt Native bilingual Fast Iteration N/A N/A Distilled: 4–8 steps 8 steps, FP8 These schemas provide the production engine with type-safe, model-specific dialects that map directly to ComfyUI node parameters while preserving cinematic intent across the distributed GPU farm.

flux2_json_schema