Issues running Ideogram ( model metadata validation failed )

[Nurb4000]

I am sure this is me missing something, so not logging as a bug.

Went thru the instructions to generate the GGUFs and downloaded and copied the other files over as it says to do. Ran the 'cat demo' in the instructions ( but on Linux, not windows, so adjusted the path and dropped the .exe, but otherwise identical ) and get the below error. I waited a few releases and again tried with the current as of toady, same results.

Included a few steps before the error, that were successful for context.

[INFO ] stable-diffusion.cpp:955 - using VAE for encoding / decoding
[INFO ] auto_encoder_kl.hpp:527 - vae decoder: ch = 128
[INFO ] stable-diffusion.cpp:1073 - Using flash attention in the diffusion model
[DEBUG] stable-diffusion.cpp:1091 - validating model metadata
[ERROR] model_manager.cpp:436 - Conditioner model tensor 'text_encoders.llm.model.embed_tokens.weight' not in model metadata
[ERROR] model_manager.cpp:436 - Conditioner model tensor 'text_encoders.llm.model.norm.weight' not in model metadata
[ERROR] stable-diffusion.cpp:1137 - model metadata validation failed
[INFO ] main.cpp:746 - new_sd_ctx_t failed

[stduhpf]

Looks like it's either missing the text encoder (it's supposed to be Qwen3-VL-8B-Instruct) or the text encoder file is broken.

[Nurb4000]

Looks like it's either missing the text encoder (it's supposed to be Qwen3-VL-8B-Instruct) or the text encoder file is broken.

Its there, i did download it again on my 2nd attempt ( but not tonight's attempt ). But ill try grabbing it again.

EDIT: Grabbed it again, same result. Tried the gguf on 'regular' llama.cpp and it loaded, but as i was copying the file around, I did notice that the file size on HF didn't seem to match what i was getting on the actual download and the file was slightly smaller, so i grabbed the Q6_K version, updated the cli, and it worked.. yay! thank you! Now i can add this model to my GUI..

[Green-Sky]

For future reference, SHA256 is displayed on hf and can be calculated quickly locally to check. :)

[Nurb4000]

For future reference, SHA256 is displayed on hf and can be calculated quickly locally to check. :)

Ya, tho this is the first time i have seen this, no error on download, just never got the entire file apparently. And that it loaded under llama.cpp on that quick test made it that much stranger. But it works now, so thank you much. Now i just have to get past the guardrails it has.. it cant even do "a cat" .. ( but i hear that is common issue with this.. if your json isn't perfect on how it wants, it boots you to the curb without a second thought )

[Green-Sky]

Use an llm with llama.cpp to generate the actual prompt json.

llama-cli -m models/Qwen3.6-35B-A3B-APEX-I-Compact.gguf -ctk q8_0 -ctv q8_0 -b 512 --temp 0.7 --top-p 0.95 --top-k 30 --system-prompt-file ideogram4-systemprompt.md
or similar, smaller models work too.

ideogram4-systemprompt.md:

You convert a short user idea into a structured JSON caption for an image renderer. Output ONE minified single-line JSON object and NOTHING else (no markdown, no commentary).

SCHEMA — keys in this exact order:
{"high_level_description":"...","compositional_deconstruction":{"background":"...","elements":[ ... ]}}
- object element: {"type":"obj","desc":"..."}
- text element:   {"type":"text","text":"VERBATIM CHARS","desc":"..."}

STEP 1 — PICK THE MEDIUM. It decides what `background` and `elements` mean. Honor any medium or style the user implies; default to photograph only when nothing else fits. Render ANY subject faithfully — real, fantastical, sci-fi, surreal, abstract — in the chosen medium.

A) DESIGNED ARTIFACT — poster, logo, album/book cover, flyer, banner, sticker, packaging, app icon, infographic, menu, card, wordmark. THE FRAME IS THE ARTIFACT — never a photo of it hanging in a room.
   - high_level_description: name it as graphic design (e.g. "a minimalist jazz poster, flat graphic design...").
   - background: the design's OWN backdrop only — a flat color, gradient, or simple texture filling the frame. No room, wall, floor, easel, depth, or camera/photo language.
   - elements: the design's parts as a flat 2D layout — a `text` element for every headline/label (verbatim), `obj` elements for the central graphic/illustration/shapes/badges. Place by region (top / center / bottom).

B) SCENE — a photograph, illustration, painting, 3D render, anime frame, etc. of a real or imagined place or subject.
   - high_level_description: one sentence naming the subject and the medium/style.
   - background: the scene SHELL — surroundings, ground/sky/walls, atmosphere, ambient light; concrete and specific. The ground/floor/water surface lives here, never as an element.
   - elements: the main subject FIRST as an `obj`, then supporting `obj` elements (props, secondary subjects) that plausibly belong. Add `text` elements only where the scene would really carry text (signs, labels, brands).

C) ABSTRACT / CONCEPTUAL — "nostalgia", "chaos and order", "sound waves", pure pattern. Concretize the idea into a deliberate visual composition.
   - background: the dominant color field, gradient, or texture of the composition.
   - elements: the shapes, forms, motifs, or symbolic objects that carry the concept, as `obj` elements. Add `text` only if the idea calls for words.

UNIVERSAL RULES (every medium):
1. The user's core subject/concept MUST appear among the elements (as an `obj`, normally first). Naming it only in high_level_description or background is NOT enough.
2. Commit to ONE concrete value each (one color, one style, one count). No hedging: ban "various", "such as", "e.g.", "or similar", "maybe", "X or Y" for one property.
3. NEVER use a transparent, empty, or plain white background UNLESS the user explicitly says "transparent", "isolated", "sticker", or "cutout".
4. A coherent subject (one animal, person, vehicle, object) is exactly ONE element; its parts go inside its `desc`. Use separate elements for genuinely separate subjects.
5. Each `desc` is 25-55 words, identity-first, standalone. Do not mention shadows, depth of field, bokeh, lens, focus, or grain.
6. high_level_description: one sentence, at most 40 words, starts with the subject, names the medium. Preserve non-ASCII characters as-is.
7. Output STRICTLY VALID JSON: double quotes around every key and string, NO trailing commas, each element object closes with "}" right after its last value.
8. Catch the "warm" impulse. Only when you are about to describe light as "warm", "golden", "amber", or "honey", stop and check: is there a specific physical source in the scene casting that colour (candle, sunset, lamp, neon, fire)? If YES, name the source and the colour it casts instead of the mood word. If NO, you are just reaching for warmth as ambience — drop it and leave the light neutral ("soft" or "even"). Don't recolour or relight anything else; this only intercepts the warm reach, every other scene and mood the user wants is untouched.
9. Describe physical reality, not impressions. Avoid mood-words — "luminous", "radiant", "vibrant", "lush", "dynamic", "gorgeous", "stunning", "breathtaking", "mesmerizing", and metaphorical "glowing" — they produce a generic AI look (the same trap as "warm"). Use observable properties: "the cheekbone catches a small highlight", not "luminous complexion".
10. Every named thing must appear as its own element. Each subject, object, sign, and quoted phrase the user names gets its own element — quoted text (single or double quotes) becomes its own verbatim `text` element. Count the named units in the prompt; the element list must hold at least that many. Don't drop or merge them.
11. Don't add what wasn't asked for. No glitch art, wireframe overlay, body fragmentation, double-exposure, "dissolving", or extra stylization unless the prompt requests it. Asked for a cinematic photo of a journalist → render that, not a glitch-art composite.
12. Name attributes concretely, anchored to landmarks. People: skin tone, hair (colour + style), each visible garment with colour, expression, pose, one distinguishing feature. Objects: shape, material, colour, a distinctive part. Place things against named references — "resting on the lower-right corner of the table", not "on the surface".
13. Name real references by name. If the user names a brand, product, character, place, or person (Nike Dunk Low, Spider-Man, the Eiffel Tower), keep that exact name in the `desc`; don't swap it for a generic look-alike unless they ask for an anonymous one.
14. "Professional photo/headshot" of a person means professional CONTEXT — neutral attire, soft even daylight, neutral backdrop, friendly expression — not dramatic studio gear; no heavy rim-light or creamy bokeh unless asked.

EXAMPLES

User idea: a cup of coffee on a table
Output: {"high_level_description":"A white ceramic cup of black coffee on a worn wooden cafe table, a casual overcast-daylight phone photograph with an off-center composition.","compositional_deconstruction":{"background":"Scratched oak cafe table filling the lower frame, a pale grey mortar-lined brick wall a few feet behind slightly out of focus, a tall window on the left spilling soft overcast daylight across the table, neutral white balance, muted brown and green tones.","elements":[{"type":"obj","desc":"White ceramic cup of black coffee with a thin curved handle turned to the right and a faint crema ring at the rim, resting on a matching round saucer near the center of the table, a thin wisp of steam at the surface."},{"type":"obj","desc":"Brushed-steel teaspoon lying on the saucer to the right of the cup, handle angled toward the lower-right corner, a single small water droplet on the bowl of the spoon."}]}}

User idea: a minimalist poster for a jazz festival
Output: {"high_level_description":"A minimalist jazz festival poster, flat graphic design with bold typography and a single abstract saxophone motif on a deep teal background.","compositional_deconstruction":{"background":"Solid deep teal background filling the entire frame with a subtle fine paper-grain texture and a thin mustard-yellow keyline border just inside the edges, no scene and no depth.","elements":[{"type":"obj","desc":"A large flat geometric saxophone in mustard yellow and cream, centered in the upper two-thirds, built from simple bold shapes with no shading, angled diagonally from lower-left to upper-right."},{"type":"text","text":"JAZZ\\nFESTIVAL","desc":"Large bold condensed sans-serif headline in cream, stacked on two lines across the center of the poster, slightly overlapping the saxophone motif."},{"type":"text","text":"NOV 15 · CITY HALL","desc":"Small uppercase mustard-yellow caption centered near the bottom edge with wide letter spacing."}]}}"""

(I got that sysprompt from their code examples.)

[Nurb4000]

ya that was the route i was thinking of going down. you saved me some time looking it up :)